Those lecture notes cover the basics of a course in microeconomic theory for MSc students in Economics. They were developed over five years of teaching MSc Economic Theory I in the School of Economics at the University of East Anglia in Norwich, UK. The lectures differ from the standard fare in their emphasis on utility theory and its alternatives.
The programme of this course is divided in three parts; choice under uncertainty, game theory and incentive theory.
The whole of the course can be covered in sixteen hours of teaching, along with eight hours of workshops, over eight weeks.
This is an intensive program that is designed both to cover the basics in each area and progress quickly to more advanced topics. The course is thus accessible to students with little background in economics, but should also challenge more advanced students who can focus on the later sections in each parts and concentrate on the suggested readings.
Exercises covering each part increase gradually in difficulty and are often of theoretical interest on their own. Detailed answers are provided.

Attribution Non-Commercial (BY-NC)

Просмотров: 1.6K

Those lecture notes cover the basics of a course in microeconomic theory for MSc students in Economics. They were developed over five years of teaching MSc Economic Theory I in the School of Economics at the University of East Anglia in Norwich, UK. The lectures differ from the standard fare in their emphasis on utility theory and its alternatives.
The programme of this course is divided in three parts; choice under uncertainty, game theory and incentive theory.
The whole of the course can be covered in sixteen hours of teaching, along with eight hours of workshops, over eight weeks.
This is an intensive program that is designed both to cover the basics in each area and progress quickly to more advanced topics. The course is thus accessible to students with little background in economics, but should also challenge more advanced students who can focus on the later sections in each parts and concentrate on the suggested readings.
Exercises covering each part increase gradually in difficulty and are often of theoretical interest on their own. Detailed answers are provided.

Attribution Non-Commercial (BY-NC)

- StockWatson_3e_EmpiricalExerciseSolutions
- Advanced Microeconomic Theory_Jehle_Reny
- Mas-Collel - Microeconomic Theory.pdf
- Lecture Notes on Microeconomic Theory - Nolan Millers (Havard)
- Kennedy a Guide to Econometrics
- Micro-economic initiatives: handbook
- Advanced Microeconomic Theory
- Microeconomic Theory by MasColell & Whinston
- Mathematics for Economists
- Microeconomics of Banking
- Helpman.krugman.1999.Market Structure and Foreign Trade
- Mas Colell Solutions
- Mas Colell Solution Manual
- Cyclopedia of Economics
- Solutions Goldberger
- Microeconomic Theory
- Jehle and Reny Solutions
- Microeconomics
- Goldberger. a Course in Econometrics
- Mas Colell Copia

Вы находитесь на странице: 1из 132

MSc students

Alexia Gaudeul∗

Abstract

Those lecture notes cover the basics of a course in microeconomic

theory for MSc students in Economics. They were developed over five

years of teaching MSc Economic Theory I in the School of Economics at

the University of East Anglia in Norwich, UK. The lectures differ from

the standard fare in their emphasis on utility theory and its alternatives.

A wide variety of exercises for every sections of the course are provided,

along with detailed answers. Credit is due to my students for ‘debugging’

this material over the years. Specific credit for some of the material is

given where appropriate.

JEL codes: A1, A23, D0.

Keywords: Microeconomics; Utility Theory; Game Theory; Incentive

Theory; Online Textbook; Lecture Notes.

is at http://agaudeul.free.fr.

1

1 PROGRAMME 2

1 Programme

The programme of this course is divided in three parts; choice under uncertainty,

game theory and incentive theory. The whole of the course can be covered in

sixteen hours of teaching, along with eight hours of workshops, over eight weeks.

This is an intensive program that is designed both to cover the basics in each

area and progress quickly to more advanced topics. The course is thus accessible

to students with little background in economics, but should also challenge more

advanced students who can focus on the later sections in each parts and concen-

trate on the suggested readings. Exercises covering each part increase gradually

in difficulty and are often of theoretical interest on their own. Detailed answers

are provided.

This course may be complemented with lectures on consumer and firm theory,

and on general equilibrium concepts. Those are not covered in those notes.

The main concepts that are covered in each part are listed below:

Rationality and axiomatic theories of choice. Expected utility theory:

its foundations and applications. Experimental tests of expected utility

theory. Alternatives to EUT: prospect theory, rank dependent expected

utility theory, cumulative prospect theory, regret theory.

A large part of this topic is not covered in standard text books and this

topic thus requires independent reading.

(or strategic)-Form Games. Dominance, Nash equilibrium, and mixed-

strategy extensions. Extensive-Form Games. Information sets, subgame-

perfect equilibria, and backward induction. Incomplete-information games,

Bayesian Games, Bayesian-Nash equilibria, forward induction. Purifica-

tion of mixed strategy equilibria. Auctions. Repeated games and the Folk

theorem.

Incentive Theory (2×2 hours lecture, 1×2 hours workshop): Principal-

agent models. Adverse selection (hidden information). Moral hazard

(hidden action). Second best and efficiency. Liability constraints. Risk

aversion. Commitment. Repeated adverse selection. Hold up problem.

Informed principal. Revelation principle. Direct Revelation Mechanism.

2 Textbooks:

The main texts for the course are:

(hereafter ‘Kreps’).

• Varian H.R., 1992, Microeconomic Analysis, Norton, 3d edition (hereafter

‘Varian’).

2 TEXTBOOKS: 3

• Mas-Colell A., M.D. Whinston and J.R. Green, 1995, Microeconomic The-

ory, Oxford University Press (hereafter ‘Mas-Colell’).

All of these cover, more or less, the entire module, except for choice under

uncertainty. Each has its strengths, though these are different. Mas-Colell

is the most up-to-date and comprehensive, but also the most advanced. It

is advised for those with a good background in microeconomics and with

good mathematical skills, and for those who wish to pursue further studies

in economics.

CONTENTS 4

Contents

1 Programme 2

2 Textbooks: 2

3 Introduction 10

4 Readings 10

4.1 Textbook Readings . . . . . . . . . . . . . . . . . . . . . . . . . 10

4.2 Other general readings . . . . . . . . . . . . . . . . . . . . . . . . 10

4.3 Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

5.1 The objects of preference and choice . . . . . . . . . . . . . . . . 11

5.2 Lotteries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5.3 The Marschak-Machina Triangle . . . . . . . . . . . . . . . . . . 14

6.1 The axioms of von Neumann and Morgenstern’s Expected Utility

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

6.1.1 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . 15

6.1.2 Transitivity . . . . . . . . . . . . . . . . . . . . . . . . . . 16

6.1.3 Continuity or Archimedean axiom . . . . . . . . . . . . . 17

6.1.4 Substitutability or Independence Axiom . . . . . . . . . . 17

6.2 A Representation Theorem . . . . . . . . . . . . . . . . . . . . . 19

7.1 The Allais Paradox and fanning out . . . . . . . . . . . . . . . . 21

7.1.1 The common ratio effect . . . . . . . . . . . . . . . . . . . 21

7.1.2 The common consequence effect. . . . . . . . . . . . . . . 23

7.2 Process violation . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

7.3 Framing effect and elicitation bias . . . . . . . . . . . . . . . . . 24

7.4 Endowment effects . . . . . . . . . . . . . . . . . . . . . . . . . . 26

7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

CONTENTS 5

8.1 Prospect Theory, Rank Dependent EUT and Cumulative Prospect

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

8.1.1 Prospect Theory . . . . . . . . . . . . . . . . . . . . . . . 27

8.1.2 Rank Dependent EUT . . . . . . . . . . . . . . . . . . . . 29

8.1.3 Cumulative Prospect Theory . . . . . . . . . . . . . . . . 29

8.2 Regret theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

II Game theory 34

9 Introduction 35

10 Readings: 35

10.1 Textbook reading: . . . . . . . . . . . . . . . . . . . . . . . . . . 35

10.2 Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

11.1 Dominance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

11.1.1 Iterated delection of strictly dominated strategies . . . . . 38

11.2 Nash equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

11.2.1 Nash equilibrium in mixed strategies . . . . . . . . . . . . 40

12.1 Backward induction . . . . . . . . . . . . . . . . . . . . . . . . . 44

13.1 Forward induction . . . . . . . . . . . . . . . . . . . . . . . . . . 46

14 Bayesian Games 48

14.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

14.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

14.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

14.3.1 Harsanyian Purification . . . . . . . . . . . . . . . . . . . 51

14.3.2 Auctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

CONTENTS 6

15 Repeated games 55

15.1 Finitely repeated games . . . . . . . . . . . . . . . . . . . . . . . 55

15.1.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

15.1.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

15.1.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

15.2 Infinitely repeated games . . . . . . . . . . . . . . . . . . . . . . 59

15.2.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

15.2.2 The Nash-Threats Folk Theorem . . . . . . . . . . . . . . 60

16 Introduction 64

17 Readings 65

17.1 Textbook Readings. . . . . . . . . . . . . . . . . . . . . . . . . . 65

17.2 Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

18.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

18.2 The first best (θ known) . . . . . . . . . . . . . . . . . . . . . . . 68

18.3 The second best . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

18.3.1 Incentive compatibility . . . . . . . . . . . . . . . . . . . . 70

18.3.2 Participation constraints . . . . . . . . . . . . . . . . . . . 71

18.3.3 The program of the principal . . . . . . . . . . . . . . . . 71

19.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

19.2 The first best outcome (perfect information on effort) . . . . . . 75

19.3 The second best outcome . . . . . . . . . . . . . . . . . . . . . . 76

19.3.1 Liability constraint . . . . . . . . . . . . . . . . . . . . . . 76

19.3.2 Risk aversion . . . . . . . . . . . . . . . . . . . . . . . . . 77

CONTENTS 7

20 Extensions 77

20.1 Repeated adverse selection. . . . . . . . . . . . . . . . . . . . . . 78

20.2 The hold up problem . . . . . . . . . . . . . . . . . . . . . . . . . 79

20.3 Other extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

20.3.1 Informed principal . . . . . . . . . . . . . . . . . . . . . . 80

20.3.2 Mixed models . . . . . . . . . . . . . . . . . . . . . . . . . 81

20.3.3 Limits to contractual complexity . . . . . . . . . . . . . . 81

20.3.4 The information structure . . . . . . . . . . . . . . . . . . 81

21.1 Application: Auctions . . . . . . . . . . . . . . . . . . . . . . . . 84

21.2 Voting mechanisms and limits to truthful DRMs . . . . . . . . . 84

IV Exercises 87

23 Game Theory 1 91

24 Game Theory 2 94

25 Incentive theory 96

LIST OF FIGURES 8

List of Figures

1 Why did the chicken cross the road? (c) Wiley Miller . . . . . . 14

2 The Marschak-Machina Triangle . . . . . . . . . . . . . . . . . . 14

3 Violation of transitivity . . . . . . . . . . . . . . . . . . . . . . . 16

4 Independence axiom . . . . . . . . . . . . . . . . . . . . . . . . . 18

5 Common ratio and common consequences . . . . . . . . . . . . . 22

6 Probability weighting function . . . . . . . . . . . . . . . . . . . 28

7 RDEUT and CPT curves . . . . . . . . . . . . . . . . . . . . . . 31

8 Best response functions in the Coordination Game . . . . . . . . 41

9 Extensive form representation of the game of entry deterrence. . 43

10 Centipede game . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

11 Coordination game with an outside option . . . . . . . . . . . . . 47

12 Convex hull for the prisoners’ dilemma . . . . . . . . . . . . . . . 60

13 Set of attainable payoffs in the prisoners’ dilemma . . . . . . . . 61

14 Poor regulatory oversight (c) Scott Adams . . . . . . . . . . . . . 65

15 The graphical derivation of the first best outcome . . . . . . . . . 70

16 The graphical derivation of the second best outcome . . . . . . . 73

17 Two lotteries and indifference curves in the Marschak-Machina

triangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

18 Extensive form game for exercise 3, Game Theory 1 . . . . . . . 92

19 Extensive form game for exercise 4, Game Theory 1 . . . . . . . 92

20 Extensive form game for exercise 5, Game Theory 1 . . . . . . . 93

21 Extensive form game for exercise 1, Game Theory 2 . . . . . . . 94

22 Range of possible MSNEs in Exercise 2, Game Theory 1 . . . . . 108

23 Range of possible MSNEs in Exercise 5, Game Theory 1 . . . . . 112

24 Extensive form game in Exercise 1, Game Theory 2 . . . . . . . . 116

25 Extensive form game in Exercise 2, Game Theory 2 . . . . . . . . 118

26 Extensive form game in Exercise 3, Game Theory 2 . . . . . . . . 120

27 Extensive form game in variant of Exercise 1, Game Theory 2 . . 121

28 First best solution in Exercise 3, Incentive Theory . . . . . . . . 124

29 Outcome of the first best contracts under asymmetric information

in Exercise 3, Incentive Theory . . . . . . . . . . . . . . . . . . . 125

30 Second best solution in Exercise 3, Incentive Theory . . . . . . . 126

31 First best solution in Exercise 4, Incentive Theory . . . . . . . . 128

32 Second best solution in Exercise 4, Incentive Theory . . . . . . . 128

9

Part I

3 INTRODUCTION 10

3 Introduction

This section will give tools to think about choice under uncertainty. A model

of behavior when faced with choices among risky lotteries will be presented

– The von Neumann-Morgenstern (“vNM”) Expected Utility Theory (“EUT”)

(1944).1 The consequences of EUT in terms of prescribed or predicted behavior

will be analyzed, and this will be compared with empirical and experimental

data on the behavior of agents. That data will be shown not to conform with

EUT in some cases, which means EUT may not be an adequate descriptive

model of behavior. Alternative models that take better account of the reality of

the patterns of decision making of economic agents will thus be presented and

discussed.

4 Readings

• Kreps, Ch. 3

• Varian, Ch. 11

• Mas-Colell, Ch. 6.

The following are useful survey papers, though note that they partly overlap:

• Machina M., 1987, Choice under uncertainty: problems solved and un-

solved, Journal of Economic Perspectives 1(1), 121-154.

• Starmer C., 2000, Developments in Non-Expected Utility Theory: The

Hunt for a Descriptive Theory of Choice Under Risk, Journal of Economic

Literature 38(2), 332-382.

The following website at The Economics New School is well designed and infor-

mative:

• Chakrabortty A., Why we buy what we buy, The Guardian, May 20th

2008.

• Do economists need brains?, The Economist, July 24th 2008.

1 von Neumann J. and O. Morgenstern, 1944, Theory of Games and Economic Behavior,

5 BASIC TOOLS AND NOTATIONS 11

4.3 Articles2

• Bateman I.J., A. Munro, B. Rhodes, C. Starmer and R. Sugden, 1997, A

Test of the Theory of Reference Dependent Preferences, Quarterly Journal

of Economics 112(2), 479-505.

• Camerer C., 1995, Individual Decision-Making, in J.H. Kagel and A. Roth,

Handbook of Experimental Economics, Princeton University Press, Ch. 8,

esp. pp. 617-65.

• Hey J. and C. Orme, 1994, Investigating Generalizations Of Expected

Utility-Theory Using Experimental-Data, Econometrica 62(6), 1291-1326.

• Kahneman D. and A. Tversky, 1979, Prospect theory: an analysis of de-

cision under risk, Econometrica 47(2), 263-291.

• Kahneman D. and A. Tversky, 1999, Choices, values and Frames, Cam-

bridge University Press.

• Machina M., 1982, Expected Utility analysis without the independence

axiom, Econometrica 50, 277-323.

• Rabin M., 2000, Risk Aversion and Expected-Utility Theory: A Calibra-

tion Theorem, Econometrica 68(5), 1281-1292.

• Shefrin H. and M. Statman, 1985, The Disposition to Sell Winners Too

Early and Ride Losers Too Long - Theory and Evidence, Journal of Fi-

nance 40(3), 777-790.

• Shogren J.F., S.Y. Shin, D.J. Hayes and J.B. Kliebenstein, 1994, Resolv-

ing differences in willingness to pay and willingness to accept, American

Economic Review 84(1), 255-270.

• Sugden R., 1985, New developments in the theory of choice under uncer-

tainty, Bulletin of Economic Research 38(1), 1-24.

• Sugden R., 2002, Alternatives to Expected Utility: Foundations, in P.J.

Hammond, S. Barberá and C. Seidl (eds.), Handbook of Utility Theory

Vol. II, Kluwer: Boston.

• Tversky A. and D. Kahneman, 1991, Loss aversion and riskless choice: a

reference dependent model, Quarterly Journal of Economics 106(4), 1039-

1061.

action are uncertain.

2 Some articles in this list were contributed by previous teachers in MSc Economic Theory

1 at the UEA.

5 BASIC TOOLS AND NOTATIONS 12

Actions can consist in buying, selling, going out, staying in, buying an um-

brella, etc.

Uncertainty may be due to the incomplete information about the action of

others, the inability to predict complex events (weather), limitations in

one’s capacity to process complex data, etc.

Outcomes are defined in terms of utility. Outcomes will be influenced by your

choice of action and the realization of random events on which you have no

influence. For example, a chicken that decides to cross a road will either

die if a car happens to come by, or live if no car happens to come by. We

assign ‘utility’ to those two possible outcomes. At its most basic, utility

is defined in terms of preferences, for example, do you prefer sun or rain?

The agent is supposed to know what acts she can choose, i.e. she knows her

options. She is also supposed to know the set of all possible ‘state of nature’

that may prevail, as well as be able to evaluate the probability, objective or

subjective, of the occurrence of each possible state of nature. She is also sup-

posed to be able to evaluate the utility of all possible consequences of each of

her actions (outcomes). Consequence functions are defined as what happens if

you chose an act and such or such state of nature occurred.

Consider for example a chicken faced with the decision whether to cross the

road. The consequence function of crossing the road combined with the event

that a car is on the road is what happens in that case, i.e. the chicken is run

over and dies. This can be denoted as follows:

c(car the road},cross

| on {z | {z road}) = death

the | {z }

event action outcome

Under this setting, you are supposed to know the utility of the outcome of

your whole set of action under any possible circumstances, e.g. the chicken

is supposed to be able to anticipate any possible event that may occur when

he crosses the road, and the outcome of his action under any of those events.

You know this in advance even though you may never have experienced such

combination of circumstances.

Obviously, your choice would be greatly simplified if a function allowed you to

calculate the expected utility of an action based on the expected probability of

the outcomes that will result from that action and an evaluation of the utility

of each of those outcomes, so as to obtain automatically your expected utility

from an action.

Why would you want to predict the utility of an action? This is because you

are constantly having to make decisions under uncertainty, and want to take

the decision that will maximize utility. We will see in the following that subject

to some assumptions on how you make basic decisions and rank the utility of

events, then one obtain a very simple way to evaluate the utility of any outcome.

Let us first present what is a lottery and how they can be represented graphically.

5.2 Lotteries

Lotteries are diagrammatic representations of the choice you are facing. Simple

lotteries assign probabilities to outcomes, which can be numerical, monetary or

5 BASIC TOOLS AND NOTATIONS 13

2. (1, ..., s, ..., S) is the set of states of nature, for example (car on the road,

no car on the road)

3. (1, ..., x, ..., X) is the set of actions that are available to you, for example

(cross the street, don’t cross the street)

4. c(s, x) is a consequence function or outcomes, which is dependend both on

the state of nature s and your action x. For example, c( car on the road,

cross the street) is ‘death’, as explained above.

5. p(s) is a probability function, that gives out the probability of each state

of nature. For example, p( car on the road) = 0.7.

6. Most of the time, one will take action x as given and denote c(s, x) in

short hand as cs and p(s) as ps . Then, the lottery L that results from

action x is defined by the vector C = (c1 , ..., cs , ..., cS ) of possible conse-

quences of x and its vector P = (p1 , ..., ps , ..., pS ) of the probabilities of

each consequences in C.

7. Most of the time, one will take the set of consequences as a given, so that

lottery L will be denoted by its vector P of probabilities associated to

each consequences.

8. Take the set of consequences C = ( death, life on this side of the road,

life on the other side of the road ). Consider L1 the lottery which results

from the action ‘cross the street’ and L2 the lottery which results from

the action ‘don’t cross the street’. In short-hand, one can denote L1 as

(0.7, 0, 0.3) and L2 as (0, 1, 0).

and this guarantees the chicken crosses the road.

(b) If the chicken is indifferent between lottery L1 and lottery L2 , one

will denote this as L1 ∼ L2 , and the chicken may or may not cross

the road.

(c) If the chicken considers the lottery L1 to be at least as good as L2 ,

one will denote this as L1 L2 and say that the chicken has a

weak preference for L1 . In that case, one does not know whether the

chicken is indifferent between L1 and L2 or if it actually prefers L1

over L2 . One knows however that the chicken does not prefer L2 over

L1 . From an observational point of view, if the chicken crosses the

road, one knows that L1 L2 but nothing more.

9. If the chicken crosses the road, then that means that L1 L2 . However,

saying that L1 L2 does not give any explanation for the behavior of

the chicken, but merely translates an observation of its behavior. I do not

know why the chicken crossed the road, I merely know it did.

5 BASIC TOOLS AND NOTATIONS 14

Figure 1: Why did the chicken cross the road? (c) Wiley Miller

Denote C = {A, B, C} a set of consequences and P = (pA , pB , pC ) a vector of

probability defined over C. Note that one will always have pA + pB + pC = 1,

as an event will always occur (‘no event’ is itself an event...). The set P of

all possible vectors of probabilities is depicted below in the Marschak-Machina

Triangle (‘M-M Triangle’).3

pB

pC

. L

pA

B C

3 For

more details, read Machina M.J., 1987, Decision Making in the Presence of Risk,

Science 236, 537-543. This expository device was introduced in Marschak J., 1950, Rational

behavior, uncertain prospects, and measurable utility, Econometrica 18, 111-141.

6 EXPECTED UTILITY THEORY 15

The corners represent the certainty cases, i.e. A is such that pA = 1, B is such

that pB = 1 and C is such that pC = 1. A particular lottery, L = (pA , pB , pC )

is represented as a point in the triangle. By convention, cA cB cC : the

top corner is preferred to the origin which is preferred to the right corner. Any

lottery over three events can be represented. pA is measured along the vertical

axis, pC is measured along the horizontal, leaving pB to be measured from the

other vertex. Thus at the origin, B, where pA and pB are both equal to 0, then

pB = 1. The hypotenuse is the range of lotteries for which pB = 0, i.e. the

hypotenuse is all lotteries that combine outcome A and C only.

We will be interested in the shape of utility indifference curves, which connect all

lotteries over which the agent is indifferent. The Marschak-Machina triangle will

allow us to show utility indifference curves in a probability space, represented by

the triangle. The representation of lotteries in the Marschak-Machina triangle,

as well as the representation of preferences over lotteries in this same M-M

triangle, will repeatedly be used to illustrate theoretical proposition introduced

in those lectures.

Exercises:

(0, 1, 0), L2 = (0.5, 0, 0.5), L3 = (1/3, 1/3, 1/3)

2. Represent the lotteries the chicken is facing when crossing the road,

assuming the ordered set of consequences is (death, life on this side

of the road, life on the other side of the road) and the probability a

car is on the road is 0.7.

3. Represent one possible utility indifference curve representing the pref-

erences of the chicken if it is found to cross the road. What properties

must the utility indifference curve have?

pected Utility Theory

von Neumann and Morgenstern (“vNM”) Expected Utility Theory (“EUT”) offers

four axioms that together will guarantee that preferences over lotteries can be

represented through a simple utility functional form.

6.1.1 Completeness

6 EXPECTED UTILITY THEORY 16

This axiom guarantees that preferences are defined over the whole set of possible

lotteries. Graphically, in the M-M triangle, this axiom guarantees that any two

points in the triangle are either on the same indifference curve or on two differ-

ent curves. Indeed, according to the axiom, either the consumer is indifferent

between L1 and L2 , in which case both lotteries are on the same indifference

curve, or they have a strict preference over L1 and L2 , in which case they are

on different indifference curves.

6.1.2 Transitivity

This axiom guarantees there are no cycles in preferences, i.e. a situation where

I prefer bananas to apple, apple to oranges and oranges to bananas is not pos-

sible... Graphically, this axiom guarantees that indifference curves in the M-M

triangle do not cross. Indeed, if there is intransitivity, then indifference curves

must cross inside the triangle. For example, below, I represent indifference

curves such that L1 L2 and L2 L3 (note that crossing outside the triangle

is not a problem). Now, you can check that if I want to represent an indifference

curve such that L3 L1 (which is intransitive), then it must cross one or the

other or both of the two previous indifference curves inside the triangle (Figure

below). Conversely, if indifference curves cross inside the triangle, then there

may be situations where transitivity does not hold.

A

L2fL 3

L fL

1 2

. €

. . fL

L3 L3 1

L1

€

L2

B €

C

6 EXPECTED UTILITY THEORY 17

there exists a unique α, 0 ≤ α ≤ 1 such that αL3 + (1 − α)L1 ∼ L2 .

cause this axiom guarantees that any point in the triangle (any lottery) has an

equivalent along the hypotenuse, and that this equivalent is unique. Suppose

indeed that L3 is the top corner of the triangle and L1 the right corner of the M-

M triangle, and consider L2 any point in the triangle. The axiom says that for

some α in [0, 1], L2 will be equivalent in terms of preferences to αL3 + (1 − α)L1

. But αL3 + (1 − α)L1 is a point on the hypotenuse of the triangle. Therefore,

any lottery has an equivalent along the hypotenuse, and conversely. This means

that there are no spaces either on the hypotenuse or in the triangle that would

have no equivalent (there are no ‘jump’ in the indifference curves).

for any α ∈ (0, 1), (1 − α)L1 + αL3 (1 − α)L2 + αL3 .

The independence axiom guarantees that indifference curves are parallel straight

lines in the M-M triangle. Indeed, represent L1 , L2 and L3 , three lotteries, in

the M-M triangle. (1 − α)L1 + αL3 and (1 − α)L2 + αL3 are parallel translations

of L1 and L2 into the M-M space, so a line that links L1 and L2 will be parallel

to a line that links (1 − α)L1 + αL3 and (1 − α)L2 + αL3 . Suppose L1 ∼ L2 then

I must have (1 − α)L1 + αL3 ∼ (1 − α)L2 + αL3 . Therefore, indifference curves

will be parallel across the whole of the triangle (it is just a matter of choosing

α and L3 ) (Figure below).

6 EXPECTED UTILITY THEORY 18

L1

L2

. (1-a)L1+aL3

.

. .

.

(1-a)L2+aL3

L3

B

C

independence means. Suppose you are in Sydney and you are offered

lotteries over the outcome space C =(train ticket to Paris, train ticket to

London). Do you prefer p = (1, 0) or q = (0, 1), i.e. a ticket to Paris or a

ticket to London? Suppose now the outcome space is C =(ticket to Paris,

ticket to London, movie about Paris). Do you prefer p0 = (0.8, 0, 0.2) or

q 0 = (0, 0.8, 0.2)?

Answer: Many students who say they prefer p to q will also say they prefer

q 0 to p0 . It may be that this type of choice is the result of improper

understanding of what is a lottery rather than something more basic, so

agents do not want p0 because “there is no point watching a movie about

Paris if I go there anyway”. Indeed, students may not understand that

each lottery will result in only one of the outcomes being realized, not two

combined outcomes together. An alternative explanation may be that

students do not want to face, under p0 , the prospect of watching a movie

about Paris which would make them regret even more not having won

the ticket to Paris. In any case, this type of reversal of preferences is a

violation of the independence axiom, as adding a common consequence to

the original lotteries should not change preferences among lotteries.

6 EXPECTED UTILITY THEORY 19

Theorem 1. [Representation theorem] If the four axioms presented above hold,

then there exists a utility index u such that ranking according to expected utility

accords with actual preference over lotteries. In other words, if we compare two

lotteries, L1 and L2 represented by the probability vectors P = (ps )s=1,...,S and

Q = (qs )s=1,...,S over the same set of outcomes S = (1, ..., S) then

s=S

X s=S

X

L1 L2 ⇔ ps u(cs ) ≤ qs u(cs )

s=1 s=1

assumes there exist a worst, w, and a best lottery, t, and defines the utility

of any lottery p as the number such that p ∼ u(p)t + (1 − u(p))w. By con-

tinuity, that number exists and is unique. p itself is defined over the set of

consequence C : (a, b, c), and each of those consequences can be ascribed

an utility according to the above method. Define thus u(a) the utility

of consequence a for example. We have p ∼ p(a)a + p(b)b + p(c)c (by re-

ducibility). We have a ∼ u(a)t+(1−u(a))w, and similarly b and c. By the

independence axiom, p ∼ (p(a)u(a)+p(b)u(b)+p(c)u(c))t+(1−p(a)u(a)−

p(b)u(b) − p(c)u(c))w. Therefore, u(p) = p(a)u(a) + p(b)u(b) + p(c)u(c).

This means the specific value u(cs ) is not what is important, rather it is

the ranking of lotteries which must be translated in a correct way by the

utility function. This means you do NOT have to translate outcomes into

one common measure, such as for example money.

• Said in another way, the theorem is a representation theorem – in other

words it means we can represent preferences using the utility function, but

that doesn’t mean that individuals gain ‘utility’ from outcomes.

• To make the point further, u(.) is unique only up to a positive, linear

transformation. This means that u(.) and any U (.) = a + bu(.) such that

b > 0 represent the same utility function. This means utility numbers

have no meaning per se.

• This utility functional has a long history dating back to Bernouilli. The

contribution of vNM was to show this was the only type of utility func-

tional that respected the above series of four normatively reasonable ax-

ioms.

• EU is linear in probabilities. vNM’s EUT makes it possible to obtain

preferences between complex lotteries through a simple adding up of the

utility of each of the components of the lottery weighted by their respective

probabilities.

• EU indifference curves in the Marschak-Machina triangle are represented

by parallel lines, sloping upward. Utility increases north-west.

6 EXPECTED UTILITY THEORY 20

Whether they fit real choice (positive axioms) or are good guides for action

(normative axioms) is up to one’s perspective. The axioms, while reason-

able, are not necessarily prescriptive or necessarily backed up or drawn

from experimental evidence. One will see indeed that agent’s actions do

not necessarily fit with EUT, which means one or more of the axioms are

not respected.

• Together, the axioms imply EUT while EU representations imply the ax-

ioms are verified.

• You need to know, understand and be able to use the axioms.

lottery L over x1 , x2 and x3 :

Rewrite this as

= (4)

U (x3 ) − U (x2 ) dp1

which is a constant (check indeed that the ratio is independent of the specific

normalization chosen for the utility functional). Therefore, dp

dp1 , which is the

3

slope of indifference lines in the M-M triangle, is always the same, no matter

where we are in the Marschak-Machina triangle. This means that once we

have found two points in the M-M triangle which give the same utility to a

particular person, then we can predict how that person will choose between any

two point in the M-M triangle. It is this remarkable feature of expected utility

theory which makes it so straightforward to test: you need only find two (non-

degenerate) indifferent lotteries to know the preferences over the whole set of

lotteries.

Checking that EUT holds means one will have to check that indifference curves

are indeed straight parallel lines, by asking agents for their preferences over

lotteries.

7 CRITIQUE OF THE AXIOMS OF EUT 21

Application to risk

Consider an agent with wealth $10,000 and utility normalized to u($x)=ln(x) de-

fined over monetary outcomes, i.e. for example, $10 provides utility u($10)=ln(10) =

2.3026. Suppose this agent is offered a lottery (1/2, 1/2) over the set of conse-

quences (-$100,$101). Will the agent accept the bet?

• If the agent prefers not betting and keeping her $10,000, she gets utility

ln(10000) = 9.2103404.

• If she bets, then she loses $100 with probability 1/2 and gains $101

with probability 1/2, so her expected utility is 1/2 ln(10000 − 100) +

1/2 ln(10000 + 101) = 9.2103399.

• This is less than ln(10000), so the agent rejects the bet.

So far, so good.

• Now, suppose the agent is offered a lottery (1/2, 1/2) over the set of conse-

quences (-$800,$869). Then you can check she rejects this as well. Indeed,

ln(10000) = 9.2103404 < 1/2 ln(10000 − 800) + 1/2 ln(10000 + 869) =

9.2103144.

• Suppose now she is offered a lottery (1/2, 1/2) over the set of consequences

(-$8000,$38476). Then she rejects this as well. Indeed, ln(10000) =

9.2103404 < 1/2 ln(10000 − 8000) + 1/2 ln(10000 + 38476) = 9.1948633.

The Allais paradox (1953)5 was first expressed in the context of the common

ratio effect, and was then generalized to include the common consequences effect.

Machina triangle and which, when connected, form a parallelogram. Let them

be defined over outcomes (consequences) x1 = $0, x2 = $50 and x3 = $100.

1. p1 is (0%, 100%, 0%): : $0 with 0% chance, $50 with 100% chance, $100

with 0% chance.

4 Rabin M., 2000, Risk Aversion and Expected-Utility Theory, Econometrica 68(5), 1281-

1292.

5 Allais M., 1953, Le comportement de l’homme rationel devant le risque, critique des

7 CRITIQUE OF THE AXIOMS OF EUT 22

2. p2 is (1%, 89%, 10%): $0 with 1% chance, $50 with 89% chance, $100 with

10% chance.

3. q1 = (89%, 11%, 0%): $0 with 89% chance, $50 with 11% chance, $100

with 0% chance.

4. q2 = (90%, 0%, 10%): : $0 with 90% chance, $50 with 0% chance, $100

with 10% chance.

When confronted with this set of lotteries, there are people who choose p1 over

p2 and choose q2 over q1 . This contradicts the independence axiom of expected

utility, as we are going to prove, both diagrammatically and analytically, that

an expected utility maximizer who prefers p1 to p2 ought to prefer q1 to q2 . This

contradiction is called the “Allais Paradox”.

Diagram (see figure below): Consider the diagram below where the four lot-

teries above are represented. Note how the line that connects p1 and p2 ,

and the one that connects q1 and q2 , are parallel. If indifference curves

are parallel to each other, as in EUT, then it should be if p1 p2 then

q1 q2 . We can see this diagrammatically by considering an indifference

curve which translates the fact that p1 p2 : it ought to separate p1 from

p2 , with p1 above and p2 below. A parallel indifference curve will divide

q1 from q2 , with q1 above and q2 below, from which one can conclude that

q1 q2 .

.

p’2

.

p2

. q2

B

.

p1 .

q1 C

7 CRITIQUE OF THE AXIOMS OF EUT 23

von Neumann-Morgenstern expected utility representation, there is some

elementary utility function u such that:

then subtracting 0.89u($50) from both sides, the first equation implies:

(8)

Combining the similar terms together, this means:

With reference to the graph above, tests of the common ratio effect involve pairs

of choices like (p1 or p2 ) and (q1 or q2 ). Tests of the common consequence effect

involve pairs of choices like (p1 or p02 ) and (q1 or q2 ).

Example: Consider the choice between p1 = (0%, 100%, 0%) and p02 = (9%, 0%, 91%).

Suppose the agent prefers p1 to p02 but prefers q2 to q1 . Graphically, one

can see that this contradicts the independence axiom too.

The ‘common consequence’ effect is less robust than the ‘common ratio’ effect,

i.e. the contradiction in choice is less often observed in that case. This may

be due to the simpler nature of common consequences lotteries, which as the

name indicates involve comparison between lotteries defined across a maximum

of two consequences only, rather than a maximum of three in the common ratio

effect. Part of the issue with the common ratio effect may thus be due to

how complicated that type of choice is, rather than to some inherent behavior

pattern.

Consider the following two lotteries:

7 CRITIQUE OF THE AXIOMS OF EUT 24

2. $-bet: $100 with 30% probability and zero otherwise.

An agent is first asked which lottery they would prefer playing, and then asked

what price they would buy a ticket to play the P-bet or the $-bet. Notice that

in this example, the expected payoff of the $-bet is higher than that of P-bet,

but the $-bet is also more risky (lower probability to win). This may be why

people tend to choose to play the P-bet over the $-bet. Yet, the same people are

ready to pay more for a ticket to play the $-bet than for a ticket to play a P-

bet. Said in another way, although when directly asked, they would choose the

P-bet, they are willing to pay a lower certainty-equivalent amount of money for

a P-bet than they do for a $-bet. For example, they might express a preference

for the P-bet, but be ready only to pay $25 to play the P-bet while being ready

to pay $27 for the $-bet.

Many have claimed that this violates the transitivity axiom. The argument is

that one must be indifferent between the certainty-equivalent amount (“price”) of

the bet and playing it, so that in utility terms, taking the example above again,

I would have U (P − bet) = U ($25) and U ($ − bet) = U ($27). By monotonicity,

since more money is better than less money, U ($27) > U ($25) and so we should

conclude that U ($ − bet) > U (P − bet). Yet, when asked directly, people usually

prefer the P − bet to the $ − bet, implying U (P − bet) > U ($ − bet). Thus, the

intransitivity.

However, the question is whether this “intransitivity” is not simply due to

overpricing of $-bets, i.e. agents being unable to price bets correctly. For

example, I once offered to a student a ticket for a bet giving 21 chance of $100

and 12 chance of $0. I was offered $12 in exchange for that ticket. When I asked

the same person how much they would sell me a ticket for this bet, they quoted

a price of $43... Was the discrepancy due to improper understanding by the

student, to lack of experience, to an exaggerated fear of getting it “wrong” in

front of others, or to the fact this was not real money? Would the discrepancy

have survived if the student had been given the opportunity to change his bids,

or if a negotiation process had been put in place? One could design a series of

alternative design for the experiment, but it illustrates a Willingness to Accept

/ Willingness to Pay disparity that survives whatever the experimental set-up

(on the discrepancy WTA/WTP, see ‘endowment effect’)

In order to determine if the P-bet, $-bet anomaly is due to the procedure by

which the preference between lotteries is elicited, or to true intransitivity in

preferences, consider an experiment mentioned by Camerer (1995).6 In this

experiment, the P-bet offers $4 with probability 35 36 , 0 else, while the $-bet

offers $16 with probability 11

36 , 0 else.

The subject is asked to choose between the two, and then is asked how much

they are prepared to pay for each of them:

6 Camerer C., 1995, Individual decision making, in The handbook of experimental eco-

nomics, J.H. Kagel and A.E. Roth (Eds), Princeton University Press, p. 660.

7 CRITIQUE OF THE AXIOMS OF EUT 25

• If the subject chose the P-bet and quoted a higher price (e.g. $3.5) for

it than for the $-bet (e.g. $3), then the preferences were judged to be

coherent.

• If the subject chose the P-bet but then quoted a lower price (e.g. $3.5)

for it than for the $-bet (e.g. $4.5), then this was judged to potentially

indicate a reversal of preference. In order to determine if that potential

reversal was merely due to overpricing of the $-bet, experimenters then

asked the subject to choose between each bet and a stated amount, $4 in

this case.

– If the subject chooses $4 rather than the $-bet, then that means they

previously overpriced the $-bet and there was thus a framing effect

at play rather than a true violation of transitivity. Indeed, the only

reason there was an apparent violation of transitivity is merely that

in one case the subject was asked to make a choice between lotteries,

and in the second case, he was asked to price lotteries. Those are

rather different mental processes.

– If on the other hand the subject indeed choose the $-bet over the $4

and the $4 over the P-bet, then that meant there was true violation

of transitivity, independent of any framing effect. Indeed, the ques-

tion was framed the same way (choice between lotteries) and led to

a contradiction of transitivity.

may occur because agents use rather less sophisticated techniques than EUT

to evaluate lotteries. They posit that agents tend to choose lotteries with the

larger probability as long as the payoffs are close, and choose lotteries with the

larger payoff is the payoffs are far away. This can lead to intransivities. For

example, suppose an agent is asked to choose between 60% probability of get-

ting $8 (L1 ) and 30% probability of getting $18 (L2 ). Payoffs are close, so most

agents choose the first as it has higher probabilities to win. Suppose then the

agent is offered $4 for sure against 60% probability of getting $8. Again, payoffs

are close, so most people agents choose the first. Finally, however, most agents

would choose 30% probability of getting $18 rather than $4 with certainty, as

payoffs are far apart so the lottery with the higher payoff is preferred. This leads

to a violation of transitivity. Indeed, the first choice implies u(L1 ) > u(L2 ). The

second choice implies u(L1 ) < u($4). The third choice implies u(L2 ) > u($4).

Combining the two last choices, u(L2 ) > u(L1 ), which contradicts the original

choice between lotteries. As above, the advantage of this design is that the

agent is always asked for their choice among lotteries and are never asked to

evaluate them individually.

This can be generalized to say that agents will try to minimize the number of

informations on which to draw their decisions rather than taking into account

all the parameters in the decision. E.g. when choosing between two brands

7 Loomes G., C. Starmer and R. Sugden, 1991, Observing Violations of Transitivity by

7 CRITIQUE OF THE AXIOMS OF EUT 26

of foods, you will not look all the ingredients that came into their making but

rather will choose salient dimensions such as sugar content and price to make

your decision. For a very interesting account of the kind of reasoning people

do when buying, see Viswanathan, Rosa and Harris (2005).8 For more on de-

signing experiments to avoid ‘framing’ subjects, i.e. to avoid obtaining results

that are dependent on the design on the experiment rather than on true subject

behavior, read Machina (1987).9

The endowment effect is such that people are less willing to pay for an object

than they would ask as payment if they owned it. This is the difference between

the WTP (Willingness To Pay) and the WTA (Willingness To Accept). This

‘endowment effects’ is not explained by wealth effects, but may be due to loss

aversion (which will come up in ‘regret theory’). For example, if losses are more

painful than equally sized gains are pleasurable, then one will offer to pay only

$12 to play a lottery with 50% chance of getting $100 so as to minimize loss in

case of bad luck (when the lottery draws $0, so the loss is $12), while one will

be prepared to sell the bet only at the much higher $43 to minimize loss in case

of bad luck, in this case, when the lottery draws $100 (the loss is $57).

Note that the disparity between WTP and WTA is explained in some measure

within EUT, but not to the extent it appears in reality. For more on this, see

exercise 5 of Choice under Uncertainty.

7.5 Discussion

Those experiments that contradict the predictions of EUT are interesting be-

cause they do not require to estimate utility functions for individuals, and they

allow for a direct test of the axioms. The drawback is that they do not allow

one to know how ‘badly’ inconsistent the choices of the agents are. This would

require many more experiments. The only information from those experiments

is the percentage of individuals whose choices violate EUT, and in what specific

way the choice is violated. An obvious critique of the above experiments is that

the choices asked from subjects in those experiments are rather too complicated

and unintuitive compared with the type of choices they are facing in real life.

The issue is then to present the problem in an intuitive form, or to allow the

subjects to get acquainted with the way the experiment is set up and the way

its consequences will affect them (i.e. let them play many rounds).

Note however that some of the experiments exposed above were adapted to

rats using food as a currency and those experiments showed violations of the

independence axiom as well. This points to a possible evolutionary benefit of

8 Viswanathan M., J.A. Rosa and J.E. Harris, 2005, Decision Making and Coping of Func-

tionally Illiterate Consumers and Some Implications for Marketing, Journal of Marketing

69(1), 15-31.

9 Machina M.J., 1987, Choice Under Uncertainty: Problems Solved and Unsolved, The

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 27

behaving in ways that are different from those of an expected utility maximizer.

Some experiments have been done comparing how farmers in traditional farming

communities evaluate lotteries, and have indeed shown their behavior, while

different from urbanized people, may have evolutionary benefits in a context

where droughts and famines are likely (Humphrey and Verschoor (2004).10

Contradictions within the framework of EUT led to the development of various

alternatives to EUT, of which we will present some. In a first part, we consider

prospect theory and its successors, in a second part we consider regret theory.

Starmer (2000)11 provides good further reading on this topic.

lative Prospect Theory

ory (‘PT’) by Kahneman and Tversky (1979),12 rank dependent utility theory

(‘RDEUT’) (Quiggin, 1982),13 and Cumulative Prospect Theory (‘CPT’) (Kah-

neman and Tversky, 1992).14 Those utility representations differ from EUT by

considering subjective probabilities as a function of objective probabilities, that

correspond to how people estimate probabilities in reality (probability weighting

functions). They also differ by considering an “editing phase”, whereby one di-

vides the set of consequences of a lottery into either ‘gains’ or ‘losses’. Different

probability weighting functions are assigned to probabilities of a consequence

depending on whether the consequence is a gain or a loss.

consider lottery L = (pa , pb , pc , pd ). Under prospect theory, its utility function

is of the form:

U (L) = π(pa )u(a) + π(pb )u(b) + π(pc )u(c) + π(pd )u(d) (10)

10 Humphrey S.J. and A. Verschoor, 2004, Decision-making Under Risk among Small Farm-

11 Starmer C., 2000, Developments in Non-expected Utility Theory: The Hunt for a De-

scriptive Theory of Choice under Risk, Journal of Economic Literature 38(2), 332-382.

12 Kahneman D. and A. Tversky, 1979, Prospect Theory: An Analysis of Decision under

13 Quiggin J.C., 1982, A theory of anticipated utility, Journal of Economic Behavior and

14 Kahneman D. and A. Tversky, 1992, Advances in prospect theory: cumulative represen-

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 28

!(p)

!(p*)=p* 1 p

The reason for offering that type of representation is that there tends to be a dif-

ference between ‘psychological probability’ and ‘objective probability’: agents

are quite able to estimate lotteries with probability between around 1/3 and

2/3 but overestimate low occurrence events and underestimate high occurrence

events. Below p∗ in the graph above, the agent weighs events above their sta-

tistical probability, while this is the opposite for events with statistical prob-

ability above p∗ . Intuitively, agents overweigh the probability of events with

low statistical probabilities, maybe because they are not used to estimating the

probability of their occurence and they feel anxious about them. The other side

of the coin is that they will under-estimate the probability of events that are

statistically quasi-given. If for example I over-estimate the probability of being

run down by a car (low probability event), then this means I underestimate the

probability that I will cross the road safely (high probability event). p∗ , which is

the probability that agents are able to estimate correctly, is variously evaluated

between 0.3 and 0.5.

Those systematic “errors” in the perception of probabilites mean that for exam-

ple, a bet offering 5% chance of $100 may sell for $6, 50% chance may sell for $50

while 95% chance may sell for $93. EUT would find it difficult to explain how

risk aversion would vary so much depending on the bet (the agent is first risk

loving, then risk neutral, then risk averse). CPT would explain this variation

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 29

π(50%)u(100) + π(50%)u(0) = u(50) (12)

π(95%)u(100) + π(5%)u(0) = u(93) (13)

π(50%) = u(50)/u(100) (15)

π(95%) = u(93)/u(100) (16)

π(50%) = 1/2 (18)

π(95%) = 93/100 (19)

CPT thus does not require one to assume anything other than risk-neutrality

to explain the pricing of the different lotteries above.

RDEUT differs from PT in that it requires that the sum of probabilies over a

whole set of events be equal to one. This is doe as follows: If one orders events

from the least to the most preferred, ( a ≺ b ≺ c ≺ d), then one writes the

utility of lottery L = (pa , pb , pc , pd ) as:

U (L) = f (pa )u(a) + f (pb )u(b) + f (pc )u(c) + f (pd )u(d) (20)

with

f (pb ) = π(pa + pb ) − π(pa ) (22)

f (pc ) = π(pa + pb + pc ) − π(pa + pb ) (23)

f (pd ) = 1 − π(pa + pb + pc ) (24)

Compared with PT, RDEUT guarantees that the sum of the probability weights

assigned to each events sum up to 1. Check indeed that f (pa ) + f (pb ) + f (pc ) +

f (pd ) = 1. In RDEUT, f (.) is a cumulative probability function, with outcomes

added in the order of their preferences.

CPT differs from RDEUT by (arbitrarily) dividing the set of consequences into

losses and gains and assigning different probability weighting function to them.

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 30

Suppose for example that I consider a and b as losses and c and d as gains.

Defining fG the probability weighting utility function applied to gains, and fL

the probability weighting function applied to losses, I then have

U (L) = fL (pa )u(a) + fL (pb )u(b) + fG (pc )u(c) + fG (pd )u(d) (25)

with

fL (pb ) = πL (pa + pb ) − πL (pa ) (27)

fG (pc ) = πG (pc + pd ) − πG (pd ) (28)

fG (pd ) = πG (pd ) (29)

with πG (.) defined over gains and πL (.) defined over losses, both increasing. In

order to guarantee the sum of probabilies is one as in RDEUT, I must have that

πL (pa + pb ) = 1 − πG (pc + pd ), that is, the cumulative probabilities of loss events

is the complement of the cumulative probabilities of gain events. CPT differs

from RDEUT by taking into account whether the event is considered as a loss

or as a gain. This is intuitively justified by saying that losses affect the agent

more than gains, and are thus overweighted. The distinction between ‘gain’

and ‘losses’ is of course subjective, and must be calibrated depending on the

individual and the situation. An alternative way to consider gains and losses is

to assign different utility functionals to gains compared to losses, whereby agents

are almost risk neutral with respect to gain (for example, a 50/50 chance to gain

$100 would be evaluated at $49), while being very risk averse with respect to

losses (for example, a 50/50 chance to lose $100 would be evaluated at −$60,

i.e. the agent is ready to pay $60 not to play the lottery).

Under CPT and RDEUT, indifference curves take the following form:

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 31

U1

U2

U3

B

C

make intuitive reasonings to justify the precise shape of the indifference curves.

Consider however the indifference curve corresponding to utility U2 . If the agent

was an EU maximizer, then its indifference curve would take the form of the

dotted line. On the hypothenuse, one considers a lottery with equal probability

between A and C, which the agent evaluates as in EUT (if one assumes p∗ = 0.5).

On the line joining B and C, one has a lottery with a better outcome with high

probability, which is going to be under-evaluated, which explains the indifference

curve is above the EU curve with the corresponding utility. However, the under-

evaluation is not so high since the probability of the best event is 0. As the

probability of that best event increases, so does the discrepancy, until a point

where the discrepancy decreases again as the lotteries involve closer outcomes.

The indifference curves are thus concave.

Exercises:

level of utility U1 and level of utility U3 .

2) How does CPT change the shape of the curve compared to RDEUT?

Regret theory differs from the above theories because it does not obtain a ‘utility

function’ but rather, a gain/loss function that evaluates by how much choosing

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 32

a lottery over another and then comparing their outcome is expected to make

one happy/unhappy. To the difference of other theories, regret theory explicitly

takes into account the opportunity cost of making a decision. Decisions are

not taken in a vacuum; making one decision precludes making another one.

The theory is exposed in Loomes and Sugden (82, 86)1516 and Sugden (85).17

Agents thus faced with alternative lotteries do not seek to maximize expected

utility but rather to minimize expected regret (or maximize expected rejoicing)

from their choice.

Formally, suppose lottery p has probabilities (p1 , .., pn ) while lottery q probabil-

ities (q1 , .., qn ) over the same finite set of outcomes x = (x1 , ..., xn ). Expected

rejoice/regret from choosing p over q is:

XX

E(r(p, q)) = pi qj r(xi , xj ) (30)

i j

qj . Lottery p will be chosen over lottery q if E(r(p, q)) is positive and q will be

chosen over p if E(r(p, q)) is negative.

Note that if r(xi , xj ) = u(xi ) − u(xj ) then E(r(p, q)) is simply the difference in

expected utility between lottery p and q and regret theory leads to exactly the

same decision as EUT.

Under regret theory:

1. r(x, y) is increasing in x so that the higher the good outcome, the higher

the rejoicing.

2. r(x, y) = −r(y, x), so that regret/rejoice is symmetric: Getting the good

outcome x rather than the bad outcome y produces the same amount

of rejoicing than the amount of regret induced by getting the symmetric

outcome. The expected rejoice at a gain is the same as the expected regret

at a same sized loss,

3. r(x, y) > r(x, z) + r(z, y) when x > z > y, that is the rejoicing increases

more than proportionately with the difference in outcome. I rejoice more

if I gain $100 rather than $0 than the sum of rejoicing if I gain $50 rather

than $0 and $100 rather than $50, even though the result is the same.

Consider for example how French people would react if their rugby team

beat Australia (pride, celebration), compared to how they would react if

France beat New Zealand (the usual...), and New Zealand beat Australia

(no one cares, at least in France, except maybe in Toulouse...).

The advantage of the regret/rejoice model is that the “indifference curves” over

lotteries derived from it can be intransitive, i.e. yield up preference reversals.

From the beginning of the course, we know that this means indifference curves

15 Loomes G. and R. Sugden, 1982, Regret Theory: An Alternative Theory of Rational

16 Loomes G. and R. Sugden, 1986, Disappointment and Dynamic Consistency in Choice

17 Sugden R., 1985, Regret, recrimination and rationality, Theory and Decision 19(1), 77-99.

8 ALTERNATIVES AND GENERALIZATIONS OF EUT 33

can cross inside the M-M triangle. Exercise 6 in ‘Choice under Uncertainty’

shows that not only does RT allow for intransitive preferences, but it also man-

dates one specific direction in which preferences can be intransitive, i.e. while it

would allow A B C A, it would not allow A ≺ B ≺ C ≺ A. This means

that a test of RT is to check that intransitiveness, when it occurs, occurs in one

direction only. Experiments tend to bear this out.

The main issue with RT is that while comparisons between two lotteries are

relatively easy under its setting, comparisons between several lotteries are much

more involved. For example, while it is easy to model the choice between (mar-

riage, no marriage), it is more difficult to model the choice between (marriage,

civil partnership, none of the above). Of course, one could divide the compari-

son between those three lotteries into three comparisons between two lotteries,

but one then loses most of the point of modeling the effect of regret on lottery

choices.

34

Part II

Game theory

The structure of this lecture as well as the notations used are drawn from

Christopher Wallace’s lectures on game theory at Oxford. Additional material

on auctions and repeated games can be found in Gibbons R., 1992, A Primer

in Game Theory, FT Prentice Hall.

9 INTRODUCTION 35

9 Introduction

This section aims to present essential tools for predicting players’ actions in a

range of strategic situations. By order of complexity, we will study strategic

form games, where players choose actions at the same time so there is no con-

ditioning of actions based on the action of others, extensive form games where

players choose actions in succession but know what action was taken by others

previously so there is no uncertainty, and finally Bayesian games where players

choose action in succession but are not certain what the other played previously.

We will also cover repeated games under full information, where players condi-

tion their action on their observation of what was played in a previous stage of

the game.

The lecture builds on the analysis of one single game, a coordination game, which

is made progressively more complex so as to introduce new concepts. Those in-

clude the concept of dominant strategy, Nash equilibrium, mixed strategy Nash

equilibrium, backward induction, subgame perfect Nash equilibrium, Bayesian

Nash equilibrium and forward induction. A few other games are introduced,

includings games of auctions (to illustrate the use of Bayesian Nash equilibrium

concepts) and the Prisoners’ dilemma (in relation to infinitely repeated games).

This lecture does not introduce many applications of game theory, as it is ex-

pected the student will encounter applications relevant to his or her area of

specialization further on in the course of his or her MSc. Rather, the lecture

aims to give good mastery of notations and techniques for solving a wide range

of games.

10 Readings:

• Kreps, Chs. 11-14

• Varian, Ch. 15

• Mas-Colell, Chs. 7-9

Among the many specialist text books which are currently available, by degree

of difficulty, one finds:

advanced)

• Gibbons R., 1992, A Primer in Game Theory, FT Prentice Hall (more

advanced)

• Fudenberg D. and J. Tirole, 1991, Game Theory, MIT Press (very ad-

vanced)

11 STRATEGIC FORM GAMES 36

10.2 Articles18

• Cho I.-K. and D. Kreps, 1987, Signalling Games and Stable Equilibria,

Quarterly Journal of Economics 102, 179-222.

• Cho I.-K. and J. Sobel, 1990, Strategic Stability and Uniqueness in Sig-

nalling Games, Journal of Economic Theory 50, 381-413.

• Geanakoplos J., 1992, Common Knowledge, The Journal of Economic

Perspectives 6, 53-82.

• Goeree J.K. and C.A. Holt, 2001, Ten Little Treasures of Game Theory

and Ten Intuitive Contradictions, American Economic Review 91(5), 1402-

1422.19

• Mailath G.J., L. Samuelson and M. Swinkels, 1993, Extensive form rea-

soning in normal form games, Econometrica 61, 273-302.

• Palacios-Huerta I., 2003, Professionals play minimax, Review Of Economic

Studies 70(2), 395-415.

• Rubinstein A., 1991, Comments on the Interpretation of Game Theory,

Econometrica 59, 909-924.

• Shaked A., 1982, Existence and Computation of Mixed Strategy Nash

Equilibrium for 3-Firms Location Problem, Journal of Industrial Eco-

nomics 31, 93-96.

• Van Damme E., 1989, Stable equilibria and forward induction, Journal of

Economic Theory 48(2), 476-496.

Definition: A strategic form game is defined by:

1. Players: The set N = 1, ..., i, ..., n of agents who play the game, for

example: Adelina and Rocco.

2. Strategies: For each i ∈ N , I define the set of strategies Si with typi-

cal element si available to agent i, for example, {Cooperate, Defect}.

3. Payoffs: Denote S = (S 1 , ..., Si , ..., Sn ) the set of all strategies avail-

able to all the players, for example ({Cooperate, Defect},{Cooperate,

Defect}).

Defect), one associates payoff ui (s) corresponding to that combination of

strategies. u = (u1 , ..., ui , ..., un ) is the set of payoffs of the game, defined

for all s in S.

18 Some articles in this list were contributed by previous teachers in MSc Economic Theory

1 at the UEA.

19 This is an article that is entertaining and easy to read about how what people play may

11 STRATEGIC FORM GAMES 37

Notation: One will denote s−i = {s1 , ..., si−1 , si+1 , ..., sn } the set of actions

taken by agents other than i in the strategy profile s and S−i = {S1 , ..., Si−1 , Si+1 , ..., Sn }

the set of strategies available to players other than i.

Example: The Prisoners’ Dilemma

1. Players N = 1, 2.

2. Strategies Si = C, D, i = 1, 2

3. Payoffs u1 (C, D) = u2 (D, C) = −6, u1 (C, C) = u1 (C, C) = −1,

u1 (D, C) = u2 (C, D) = 0 and u1 (D, D) = u2 (D, D) = −4.

C D

C -1,-1 -6,0

D 0,-6 -4,-4

11.1 Dominance

equilibrium. In a strictly dominant strategy equilibrium, players play their

strategy irrespective of the action of others.

i if ui (si , s−i ) > ui (s0i , s−i ) for all s−i in S−i .

If si strictly dominates another strategies s0i , then that strategy is strictly dom-

inated and can be eliminated from consideration by i.

dominates it.

that si is by definition unique.

in Si .

dominant strategy. Note that from the above, a strictly dominant strategy is

necessarily unique.

ui (si , s−i ) for all players i ∈ N and for all strategy profiles s−i ∈ S−i , that

is, if all elements of s∗ are strictly dominant strategies.

11 STRATEGIC FORM GAMES 38

is {D,D}. Indeed, D is a strictly dominant strategy for player 1 as u1 (D, C) >

u1 (C, C) and u1 (D, D) > u1 (C, D). The same holds for player 2.

Remark: One can also define weak dominance as follows: Strategy si ∈ Si

weakly dominates (or simply “dominates”) strategy s0i 6= si in Si for player

i if ui (si , s−i ) ≥ ui (s0i , s−i ) for all s−i in S−i .

A strategy s0i that is strictly dominated will not be part of any equilibrium of

the game. One can therefore exclude strictly dominated strategies from the set

of available strategies.

L M R

U 2,3 4,1 -1,6

D 5,3 4,-5 7,4

can thus be eliminated. U is only weakly dominated by D, as it obtains

the same payoff as D when 2 plays M. It cannot thus be eliminated from

the set of available strategies.

One obtains the following remaining game:

R

U -1,6

D 7,4

strategy equilibrium of the game, obtained by iterated deletion of strictly

dominated strategies, is thus (D, R).

Indeed, in so far as players’ actions are not dependent on others’ action, those

games cannot properly be called strategic. We introduce in the following part a

concept, that of Nash equilibrium, which is of interest in proper strategic games.

Definition: A Nash equilibrium (‘NE’) is a strategy profile s∗ such that for

every player i, ui (s∗i , s∗−i ) ≥ ui (si , s∗−i ) for all si ∈ Si .

Note how different the NE concept is from the dominant strategy concept: the

NE concept does not require that s∗i be dominant for all s−i in S−i , but only

for s∗−i . That is, taking s∗−i as given, agent i must not strictly prefer to deviate

to an action other than s∗i .

11 STRATEGIC FORM GAMES 39

2. Strategies Si =Coffee (C), Pub (P), i =F,B

3. Payoffs uF (C, P ) = uB (C, P ) = 1, uF (C, C) = uB (P, P ) = 4,

uF (P, C) = uB (P, C) = 0 and uF (P, P ) = u2 (C, C) = 3.

B

C P

F C 4,3 1,1

P 0,0 3,4

However, suppose that F plays C. Then B is better off playing C. Sup-

pose that F plays P. Then B is better off playing P. Formally, uF (C, C) ≥

uF (P, C) and uB (C, C) ≥ uB (C, P ). Similarly, uF (P, P ) ≥ uF (C, P ) and

uB (P, P ) ≥ uB (P, C). This means that both {C, C} and {P, P } are Nash

equilibria of the game.

Note how we found two Nash equilibria of the game above. This generalizes to

saying that Nash equilibria are not necessarily unique, unlike dominant strategy

equilibria. However, from the definition, one can check that any dominant strat-

egy equilibrium is also a Nash equilibrium. There is no way in the game above

to choose which of the Nash equilibria is more likely to be chosen. However, the

Nash equilibrium concept allows one to say that players will play either one or

the other Nash equilibria. The Nash equilibria we found are such that players

choose actions in a deterministic way, that is, if for example the Nash equilib-

rium is {C,C}, then both players play C. Those are called pure strategy Nash

equilibria (‘PSNE’). We will see below there exists a third Nash equilibrium

of this game, where players choose actions at random according to pre-defined

probability. Those are called mixed stategy Nash equilibria (‘MSNE’).

A common way to find Nash equilibria is to use the concept of Best Response

Function, which is particularly useful when players’ action set is continuous

(such as quantity or price in a game of competition).

such that Bi (s−i ) = {si |ui (si , s−i ) ≥ ui (s0i , s−i ) for all s0i } . The BRF states

what is the best action for i the whole range of possible profile of actions

of other agents.

Definition: s∗ is a Nash equilibrium if and only if s∗i ∈ Bi (s∗−i ) for all i. This

means that elements of s∗ must be best responses to each other.

Example: In the example above, BF (C) = C and BF (P ) = P. Similarly,

BB (C) = C and BB (P ) = P.

11 STRATEGIC FORM GAMES 40

random according to pre-defined probabilities. For the sake of easy modeling, we

assume agents have access to a randomizing device (such as a coin for example),

which allows them to randomize over their actions. For example, an agent who

decides to play C with probability 1/8 in the game above can do so by saying

he will play C whenever three throws of the coin all give out ‘tail’. We will see

later whether agents can randomize in an accurate and rational way without

access to such a randomizing device.

We did not consider above the case of mixed strategies. Consider thus

strategies si of the form: i plays C with probability pi and play P with

probability 1 − pi . Strategy si can be denoted in short as pi . Suppose

players F and B play strategies pF and pB respectively. The payoff to

F of playing C is then uF (C, pB ) = 4pB + 1(1 − pB ) = 3pB + 1. Sim-

ilarly, uF (P, pB ) = 3 − 3pB . I also obtain that uB (pF , C) = 3pF and

uB (pF , P ) = pF + 4(1 − pF ) = 4 − 3pF .

Therefore, F will play C whenever 3pB + 1 > 3 − 3pB , that is, whenever

pB > 13 . She will play P whenever pB < 13 and will be indifferent between

the two actions whenever pB = 13 . Similarly, B will play C whenever

3pF > 4 − 3pF , that is if pF > 23 . He will play P whenever pF < 32 , and

will be indifferent between the two actions whenever pF = 23 .

One thus has three Nash equilibria: {pF = 0, pB = 0}and {p F = 1, pB = 1}

as before, and a mixed strategy Nash Equilibrium 23 , 31 whereby each

player plays its own favorite option with probability 23 , and the other op-

tion with probability 13 . Under that MSNE, the payoff for F is 2 and the

payoff for B is 2 as well.

Graphically, one can represent the best response functions of both players

as follows:

12 EXTENSIVE FORM GAMES WITH PERFECT INFORMATION 41

BF

BB

pB

0 1

pF

The graph can be read as follows: for any pF < 23 , B prefers to play P,

so pB = 0. For pF = 13 , B is indifferent between P and C, so pB can

be anywhere between 0 and 1. For pF > 23 , then B prefers to play C, so

pB = 1.

You can check that all Nash equilibria are at crossing points of the best response

functions, as implied by the definition of Nash equilibria in terms of BRF.

Palacios-Huerta (2003)20 is a good introduction to the debate over whether

agents are actually able to play mixed strategies. This is particularly important

when the only NE is in mixed strategies, as in the game under study in the

article (penalty strikes). Most of the evidence in experimental settings is not

encouraging (agents find it difficult to choose actions randomly). However,

Palacios-Huerta (2003) shows that professional footballers, who have sufficient

experience and incentives, appear to be able to use mixed strategies in their

choice of which side of the goal to strike a ball.

tion

We consider in the following settings in which agents choose actions in succesion

and observe what the other players played before choosing their own action.

20 Palacios-Huerta I., 2003, Professionals play minimax, Review Of Economic Studies 70(2),

395-415.

12 EXTENSIVE FORM GAMES WITH PERFECT INFORMATION 42

2. Histories: A set of histories H = (H1 , ..., Hn ) with typical member

h = (h1 , ..., hn ). h is a sequence of actions by individual players, with

typical member hi which denotes all actions taken by i in the past.

The start of the game is denoted as ∅ ∈ H is . If h ∈ H, but there is no

(h, a) ∈ H where a is an action for some player, then h is a terminal

history. That means there is no further action to be taken by any

of the players; the game is finished. The set of terminal histories is

denoted as Z ⊂ H.

3. Player function: A function P : H\Z → N, assigning a player to

each non-terminal history. This player function indicates when is the

turn of each player to play.

4. Payoffs: vNM payoffs for each i ∈ N are defined over terminal histo-

ries, ui : Z → R. This assigns a payoff to each terminal histories of

the game.

To illustrate the above, suppose one plays a prisoners’ dilemma game twice. h1

at stage 1 may be (C), that is, player 1 played C in the first stage. h1 at stage

2 may be (C,C), that is, player 1 played C in stage 1 and in stage 2. The set of

terminal histories, over which payoffs will be defined, is:

{(C,C),(C,C)},{(C,C),(C,D)},{(C,C),(D,D)},{(C,C),(D,C)},

{(C,D),(C,C)},{(C,D),(C,D)},{(C,D),(D,D)},{(C,D),(D,C)},

{(D,C),(C,C)},{(D,C),(C,D)},{(D,C),(D,D)},{(D,C),(D,C)},

{(D,D),(C,C)},{(D,D),(C,D)},{(D,D),(D,D)},{(D,D),(D,C)}. For example, in

the last history, player 1 played D in both stages, while player 2 played D in the

first stage and C in the second.

2. Histories: H : {(∅), (Stay out), (Enter), (Enter, Fight), (Enter, Accomodate)} .

The set of terminal histories is Z : {(Stay out), (Enter, Fight), (Enter, Accomodate)}

3. Player function: P (∅) = E and P (Enter) = I. This means E plays

first, and I may play only if E enters, otherwise, I does not have

anything to do.

4. Payoffs uI (Stay out) = 2, uE (Stay out) = 0, uI (Enter, F ight) =

uE (Enter, F ight) = −1, uI (Enter, Accomodate) = uE (Enter, Accomodate) =

1. Note how payoffs are defined over all possible histories/

12 EXTENSIVE FORM GAMES WITH PERFECT INFORMATION 43

Enter Accomodate

E I 1,1

0,2 -1,-1

rid of the information about histories and player functions.

Strategies S : (SE , SI ) are SI : {Fight, Accomodate} and SE : {Stay out, Enter} .

Payoffs are defined over the terminal histories that result from a spe-

cific set of strategies, so I have to define equivalence, such that u(Stay

out, Fight)=u(Stay out), u(Stay out, Accomodate)=u(Stay out), u(Enter,

Fight)=u(Enter, Fight), u(Enter, Accomodate)=u(Enter, Accomodate).

One thus obtains the following strategic form representation of the game:

Fight Accomodate

Stay out 0,2 0,2

Enter -1,-1 1,1

Once the game has been put into its strategic form, it is easy to find its

pure strategy Nash Equilibria: {Stay out, Fight} and {Enter, Accomodate}.

In order to find its mixed Nash equilibria, denote pE the probability for

the entrant to stay out, and pI the probability for the incumbent to fight.

Then

• uE (Stay out,pI ) = 0,

• uE (Enter,pI ) = −pI + (1 − pI ) = 1 − 2pI

• uI (pE ,Fight) = 2pE − (1 − pE ) = 3pE − 1

• uI (pE ,Accomodate) = 2pE + (1 − pE ) = pE + 1.

12 EXTENSIVE FORM GAMES WITH PERFECT INFORMATION 44

From those payoffs, the entrant enters whenever pI < 12 , stays out when-

ever pI > 12 and is indifferent between entering or not when pI = 21 . The

incumbent fights whenever 3pE − 1 > pE + 1, which never happens, is

indifferent between fighting or not when pE = 1, and accomodates for any

pE < 1.

There are therefore an infinity of Nash equilibria of the entry game, of

which one PSNE, {Enter, pI = 0} and an infinite

range of mixed strat-

egy Nash equilibria denoted Stay out,pI ≥ 12 . Note that the PSNE

{Stay out, pI = 1} is simply an extreme form of a MSNE where the in-

cumbent fights with probability 1.

As in the above, the usual way to find equilibria of an extensive form game is

to reduce it to a normal form game and find the NE of that normal form game.

The equilibria of the extensive form game will be a subset of the NE of the

normal form game, as the normal form game neglects important information

about the succesion of actions in the game. We see below how the set of Nash

equilibria can be parsed down through a process of backward induction.

From the above, the entrant stays out because she expects the incumbent to

fight upon entry with probability at least half. However, since the entrant stays

out, this is a completely arbitrary belief as there is no way for the entrant to

verify that prediction. Is that arbitrary belief reasonable? When the entrant

enters, then the incumbent will always find it best to accomodate. Therefore,

the only credible belief for the entrant is that when the entrant enters, then the

incumbent would accomodate. Therefore, it is not reasonable for the incumbent

to believe that the incumbent would fight with any positive probability if she

were to enter, because it is not credible for the incumbent to actually fight upon

entry. One can thus state that the only credible equilibria of the game are those

that survive a process of backward induction where, beginning from the last

stage, actions that players will not choose in that stage are eliminated from

consideration.

For example, in the game of entry deterrence, the incumbent will not fight

when called upon to choose its action. Eliminating ‘fight upon entry’ from

consideration, the game that remains is thus a choice for the entrant between

staying out, which gives payoff 0, and entering, which, by backward induction,

will give a payoff of 1. The entrant thus chooses to enter. The only equilibrium

that survives this process of backward induction is {Enter, Accomodate} .

ward induction if it does not involve any non-credible threat, that is, if all

strategies played at each stage (histories) in the game are best-response

to each other.

For example, ‘fight’ is not best response to ‘enter’, so that any equilibrium that

involves ‘fight’ in response to ‘entry’ does not survive backward induction.

13 EXTENSIVE FORM GAMES WITH IMPERFECT INFORMATION 45

1 2 1 2 1

2,5

b) How would you play the game if you were not sure the other player

follows a process of backward induction?

mation

Definition: An extensive form game with imperfect information consists of:

1. Players: As before

2. Histories: As before

3. Player function: As before

4. Payoffs: As before

5. Information: A player’s information set defines what the player knows

about the previous history of the game when he is called upon to play.

13 EXTENSIVE FORM GAMES WITH IMPERFECT INFORMATION 46

Definition: A singleton information set is such that a player knows the history

of the game that led to the stage he/she is at. A subgame is a game that

follows a singleton information set. A game can have many subgames, and

a subgame can have subgames as well.

Example: In the Prisoners’ Dilemma, no one of the players knows what the

other player played so there is no subgame over than the whole game which

follows from history H = ∅. In the entry deterrence game, the incumbent,

once it is its turn to play, knows the entrant entered. Therefore, the game

after entry where the incumbent has to choose whether to enter or not is

a subgame. The entry deterrence game has two subgames: the one that

follows from history {∅} and the one that follows from history {Enter}.

Definition: A Nash equilibrium is subgame-perfect (‘SPNE’) if the players’

strategies constitute a Nash equilibrium in every subgame.

dilemma is also its NE. The subgame perfect Nash equilibrium in the en-

try deterrence game is {Enter, Accomodate}. The outcome of the SPNE in

the centipede game is {Stop,Stop} (based on the SPNE {(Stop,Stop,Stop),(Stop, Stop)}).

In all those cases, the SPNE coincides with the NE obtained by backward

induction, either because the game is one-stage only (Prisoners’ Dilemma),

or because there is perfect information, i.e. each decision node is a single-

ton information set, that is, players know with certainty the history that

led to the current stage in the game when it is their turn to play.

From the above examples, one can say that the set of NE that survive backward

induction and the set of subgame perfect NE coincide in extensive form games

with perfect information.

The following example will allow us to introduce the process of forward induc-

tion, which is a process of elimination of candidate SPNE to further restrict the

set of reasonable NE beyond the capabilities of the process of backward induc-

tion. A SPNE that is robust to the process of backward induction may not be

robust to the process of forward induction.

ordination game. The French player has the option to stay home, where

she gets payoff 3.5 from eating a madeleine while reading a novel. The

English player gets payoff 0 in that case. If the French player goes out,

then the British player knows this (for example he calls her home and does

not get an answer). However, he does not know if she went to the pub or

to the café.

13 EXTENSIVE FORM GAMES WITH IMPERFECT INFORMATION 47

Home

F 3.5,0

Out

F

Café Pub

B B

Note that the payoff for F of staying at home makes her willing to go out

only if her preferred PSNE obtains, but not anytime else (when the other

PSNE is played, or when the MSNE is played). Intuitively, this means

she will go out only if she believes the British player will go to the café,

in which case she goes to the café, so the British player should go to the

café if he realizes she went out. We are going to see if this intuition is

borne out in the following process of finding out reasonable equilibria of

the game:

This game that has three subgames, the one starting at {∅} (the whole

game), the one starting at {Home} and the one starting at {Out}. As

seen before, there are two PSNE of the subgame starting at {Out} as well

as one MSNE. Consider thus all possible SPNE candidates once only NE

of the subgame starting at {Out} are retained.

1. {(Home, Caf é), Caf é}

2. {(Home, P ub), P ub}

3. (Home, 32 ), 13

5. {(Out, Caf é), Caf é}

6. (Out, 23 ), 13

Note that, for example in the first case, while F stays home, we still

express what would happen if she did not. Backward induction eliminates

14 BAYESIAN GAMES 48

the first potential equilibrium: if F expects {Caf é, Caf é} to be played

in the subgame starting at {Out} , then she would go out. In the same

way, the fourth and the sixth equilibrium can be eliminated as well: if she

expects {Pub,Pub} to be played in the subgame starting at Out, then she

is better off staying in. 2, 3 and 5 are thus the only SPNE of the game.

In order to select among those, one will have to refine the concept of

a SPNE by using a process of forward induction: the second equilibrium

would require B to think that when F goes out, she goes to the pub, which

would not be rational of F since she would have done better staying home

in that case. Similarly, the third equilibrium requires B to think F plays

a mixed strategy that obtains payoff of 2, even though in that case she

would have done better staying at home. The only equilibrium that thus

survives forward induction is {(Out, Caf é), Caf é}. In that equilibrium,

F does not wish she could change any of her action at any stage of the

game (i.e. once at the café, she is happy to have gone out). This confirms

our intuition as stated at the beginning of the analysis of this game.

starts from the end of the game and assumes that behavior at any stage

of the game will be rational, forward induction starts from the beginning

of the game and assumes that present behavior is optimal given what

is expected to happen later. An equilibrium will survive the process of

backward induction when it is such that the behavior of a player proves

to have been optimal given the succeding strategies played in the game.

Remark: The ‘Intuitive Criterion’ of Cho and Kreps (1987)21 uses forward

induction to restrict the set of NE by eliminating those that are based

on unreasonable out-of-equilibrium beliefs, i.e. those out-of-equilibrium

beliefs that imply a player would play against his or her own interest if

players deviated from the equilibrium. The problem with this approach is

that, since actions that are out of equilibrium are not played, there is in

principle nothing that could restrict the beliefs of someone once they are

played. In fact, it may be rational to believe that a player that plays out

of equilibrium is actually irrational, and thus hold out any beliefs about

how he might play or what are his beliefs!

14 Bayesian Games

14.1 Example

A Bayesian game differs from a game with imperfect information in that, while

in games with imperfect information the players may not know what was played

before, in Bayesian games players may not know who they are playing against.

In practice, this means they will not know what the other player will play in

21 Cho I-K. and D.M. Kreps, 1987, Signaling games and stable equilibria, Quarterly Journal

14 BAYESIAN GAMES 49

advance of the game, which is quite different from not being perfectly informed

of what is played during the game. Before going on to define further what is a

Bayesian game, let us study a specific Bayesian game, extended from our usual

Coordination Game:

Example: The British person likes the French person with probability 12 , and

dislikes her with probability 12 . The British person knows his own type,

Bl (likes) or Bh (hates). The British person is very private and does not

display his feelings openly, so the French person does not know the type

of the British person and assigns probability 12 to the British person liking

her.

If B likes F, then payoffs are as follows:

B

C P

F C 4, 3 1, 1

P 0, 0 3, 4

B

C P

F C 4, 0 1, 4

P 0, 3 3, 1

• Players: i ∈ N = {F, Bl , Bh }

• For each i, si ∈ {C, P }

Payoffs can be shown as follows, with the first payoff the payoff of the

French person, the second the payoff of the British person of type l and

the third payoff the payoff of the British person of type h. Columns show

all possible combinations of actions that B of type l or h might take.

B

Cl , Ch Cl , Ph Pl , Ch Pl , Ph

5 5

F C 4, 3, 0 2 , 3, 4 2 , 1, 0 1, 1, 4

3 3

P 0, 0, 3 2 , 0, 1 2 , 4, 3 3, 4, 1

and B can choose either C whether it is of type l or h, or C if it is of type

l and P if it is of type h, or P if it is of type l and C if it is of type h, or P

if it is of type l and P if it is of type h. In each combination of strategies,

the expected payoff of F is first, then the payoff for Bl is second and the

payoff for Bh is third.

By underlining best responses, one finds one pure strategy Nash Equilib-

rium, {C, Cl , Ph } . There also are Mixed Strategy Nash Equilibria: Sup-

pose F mixes between C and P with probability p and 1 − p respectively,

while Bl mixes between C and P with probability pl and 1 − pl respec-

tively, and Bh mixes between C and P with probability ph and 1 − ph

respectively. Under those conditions,

14 BAYESIAN GAMES 50

pl )ph + 1(1 − pl )(1 − ph )

• the expected payoff for F in playing P is 0pl ph + 32 pl (1 − ph ) + 32 (1 −

pl )ph + 3(1 − pl )(1 − ph )

• the expected payoff for Bl in playing C is 3p + 0(1 − p)

• the expected payoff for Bl in playing P is 1p + 4(1 − p)

• the expected payoff for Bh in playing C is 0p + 3(1 − p)

• the expected payoff for Bh in playing P is 4p + 1(1 − p)

2 1

Bl thus plays C s.t. p > 3, Bh plays C s.t. p < 3. There are thus five

situations:

• If p < 13 , then Bh plays C and Bl plays P, but then F is better off

playing C : this is not a Nash equilibrium.

• If p = 13 , then Bh is indifferent between playing C or P while Bl

will play P. This is an equilibrium only if Bh ’s strategy makes F

indifferent between playing C or P, so I must have ph = 32 .

• If 13 < p < 23 , then Bh plays P and Bl plays P. Then F would play

P, but then Bh would play C. This is not a Nash equilibrium.

• If p = 32 , then Bh plays P and Bl is indifferent between C and P. This

is an equilibrium only if Bl ’s strategy makes F indifferent between

playing C or P, so I must have pl = 32 .

• If p > 23 , then Bh plays P and Bl plays C. Then F plays C, which is

a Nash equilibrium.

There are thus two MSNEs, 31 , P, 23 and 23 , 23 , P in addition to the

PSNE {C, Cl , Ph } .

14.2 Definition

In the following, we provide a formal definition of a Bayesian game.

2. Types, for each player i, there is a set Ti of possibles types, labeled

ti . Player i might be of different types that have different payoffs, but

all possible types of player i play the same role in the game.

3. Actions, for each player i, there is a set Ai of possible actions, labeled

ai . This means that whatever i’s type, he has access to the same range

of action.

4. Beliefs, for each pair (i, ti ), a probability distribution over T−i , writ-

ten pi (t−i |ti ).This denotes what player iof type ti believes the type

of other player are.

5. Payoffs, for each player i, a vNM utility function ui : A × T → R.

As said above, payoffs depend not only on the identity of the player,

but on his type.

14 BAYESIAN GAMES 51

A Bayesian game can thus be written Γ : {N, {Ti }i , {Ai }i , {pi }i , {ui }i }

Remark: Beliefs are the beliefs of each player of each different type over the

possible types of each other player. In the previous game, for example,

F believed B could be of two types, l or h, with equal probability. Since

there was only one type of F, one did not need to specify different beliefs

for different types of F. Also, both Bl and Bh had the same belief about

the type of F , since F could be of only one type.

Bayesian game of incomplete information according to the notations ex-

posed above.

14.3 Examples

In the following, we consider two Bayesian games and find their Nash Equilibria.

The first example is designed to show that if players differ in their types, then

even though each type of players play pure strategies, what is observed by an

external observer are mixed strategies. The second example considers auction

mechanisms, where we show that auctions can be used as a tool to get players

to reveal their own type.

In this part, we consider how MSNEs may arise when there is some uncertainty

over the type of the agents playing the game. Consider the following Noisy

Coordination Game which is a modification of the Coordination Game: The

French person is not sure how much the British person likes going to the pub,

and the British person is not sure how much the French person likes going to

the café. Payoffs are as follows:

B

C P

F C 4 + f, 3 1 + f, 1 + b

P 0, 0 3, 4 + b

probability distribution over (0, a) : U [0, a]. Similarly, B is not sure about c,

and believes it is distributed according to the uniform probability distribution

over (0, a) : U [0, a]. Both F and B know their own type, f and b respectively.

Consider the following Bayesian strategy, such that B plays P s.t. b > b∗ , and

F plays C s.t. f > f ∗ . This is a very simple kind of threshold strategy whereby

depending on their type, players will decide either to play C or B. Note that

this is a considerable simplification from for example assuming that B would

vary the probability with which he would play C or B depending on his type.

Here, he plays pure strategies, either P or C. ∗However, from the perspective of

B, the probability that F plays C is then a−f a while from the perspective of F ,

14 BAYESIAN GAMES 52

∗

the probability that B plays P is then a−ba . Therefore, while players play pure

strategies depending on their type, this is observationaly equivalent to mixed

strategies from the point of view of the other player.

b∗ b∗

• The best strategy for F of type f is to play C s.t. a (4 + f ) + (1 − a )(1 +

∗ ∗

f ) > (1 − ba )3, so F plays C s.t. f > 2 − 6 ba .

f∗

• Similarly, the best strategy for B of type b is to play C s.t. (1 − a )3 >

∗ ∗ ∗

(1 − fa )(1 + b) + fa (4 + b), so B plays C s.t. b < 2 − 6 fa .

which leads to the following system of equations:

b∗

f∗ = 2 − 6

a

∗

f

b∗ = 2 − 6

a

which is solved for f ∗ = b∗ = 2a

6+a .

• From the point of view of F, the probability that B goes to the pub is

then

2a 2

Pr(b ≥ )=1−

6+a 6+a

• From the point of view of B, the probability that F goes to the café is

2a 2

Pr(f ≥ )=1−

6+a 6+a

One will notice that as a → 0 (uncertainty disappears), then each player can

expect the other player to play her preferred choice with probability 32 , which

corresponds to the MSNE of the game. This means that ‘purification’ of the

pure-strategy Bayesian equilibrium of this game resembles the MSNE of the

unperturbed game. Note however that the PSNE of the unperturbed game are

also Bayesian equilibria of the perturbed game. This means that purification

of pure-strategy Nash equilibria does not exclude any NE of the original game

with no uncertainty.

14.3.2 Auctions

Auctions are a specific example of Bayesian games, whereby one or many buyers

and one or many sellers must reach an agreement over the exchange of one or

many goods. This agreement may be about exchange price, quantity exchanged,

and the identity and allocation to sellers and buyers.

We will examine two very simple mechanisms:

1. One is such that a seller must decide to which of two buyers to sell one

indivisible good in its possession, and at what price.

14 BAYESIAN GAMES 53

2. The other is such that one buyer is faced with one seller with one indivisible

good, and must reach an agreement over the price at which that one

indivisible good will be exchanged.

the two situations, and there are also a variety of other situations where auction

mechanisms could be applied. We will examine only the first price sealed bid

auction as applied to the first situation, and the double auction as applied to

the second situation.

Two buyers, i and j, submit sealed bids to a seller. The highest bidder wins

the object, and pays his/her bid to the seller. The object is worth vi to buyer i

and vj to buyer j. i knows vi but not vj , and j knows vj but not vi . i believes

vj is distributed uniformly over [0, 1], j believes vi is distributed uniformly over

[0, 1].

Strategies are of the form bi : [0, 1] → R, which for any vi assigns a bid bi (vi ).

Payoffs can be represented as follows:

0 if bi < bj

vi −bi

ui (bi ; vi ) = 2 if bi = bj

vi − bi if bi > bj

We are going to looks for a linear bidding strategy, s.t. bi (vi ) = α + βvi .

Neglecting the case where bi = bj (an event of probability measure 0), and

assuming bj ∈ [α, α + β], the expected payoff to player j who is bidding bj is

then

bj − α

= Pr( ≥ vi )(vj − bj )

β

bj − α

= (vj − bj )

β

vj +α

This is a concave function which attains a maximum for bj = 2 .

vj +α

Suppose j indeed chooses bj = 2 . Then

vj + α

E(ui (bi ; vi )) = Pr(bi ≥ )(vi − bi )

2

= Pr(2bi − α ≥ vj )(vi − bi )

= (2bi − α)(vi − bi )

This is a concave function that attains its maximum for bi = α+2v4 . Combining

i

1

the two equations for bi and bj , one obtains α = 0 and β = 2 . Each bidder thus

bids half his/her own valuation for the good.

14 BAYESIAN GAMES 54

Remark: This linear bidding strategy is the unique symmetric Bayesian equi-

librium if valuations are uniformly distributed. There may be other, non-

symmetric Bayesian strategies, and if valuations are not uniformly dis-

tributed, then the linear bidding strategy is not necessarily optimal.

Exercise: Consider the three following alternative bidding rules, and solve for

the optimal bidding strategy:

• The second price auction, whereby the highest bidder wins the good and

pays the bid of the second highest bidder.

• The ‘all pay’ auction, whereby the highest bidder wins the good and both

bidders pay their own bid.

• The ‘loser pays’ auction, whereby the highest bidder wins the good and

the second highest bidder pays the bid of the highest bidder.

2) Double Auction

A buyer with valuation vb ∼ U [0, 1] for a good is faced with a seller with

valuation vs ∼ U [0, 1] for a good. Both have to submit their bids (offered price

for the buyer, asked price for the seller), simultaneously. If pb < ps then there

is no trade. If pb ≥ ps , then they split the difference, so the trade price is

p = ps +p

2 .

b

(

0 if pb < ps

ub (pb ; vb ) = ps +pb

vb − 2 if pb ≥ ps

(

0 if pb < ps

us (ps ; vs ) = ps +pb

2 − vs if pb ≥ ps

As before, one will consider linear bidding strategies s.t. pb (vb ) = α + βvb and

ps (vs ) = γ + δvs .

Consider the strategy of the buyer. The expected payoff to the buyer who is

offering pb is

E(ps | pb ≥ ps ) + pb

E(ub (pb ; vb )) = Pr(pb ≥ ps )(vb − )

2

with

pb − γ

= γ + δE(vs | ≥ vs )

δ

pb − γ

=γ+δ

2δ

1 1

= γ + pb

2 2

15 REPEATED GAMES 55

so

1

ps − γ γ + 12 pb + pb

E(ub (pb ; vb )) = Pr( ≥ vs )(vb − 2 )

δ 2

pb − γ γ + 3pb

= (vb − )

δ 4

This is a concave function that is maximized for pb = γ+2v 3 . Similarly, the

b

α+β+2vs

optimal bidding strategy for the seller is to ask ps = 3 . Combining the

two equations for pb and ps with the assumed expression of their value, one

obtains β = δ = 23 . while α = γ3 and γ = α+β 1

3 . This is solved for γ = 4 and

1

α = 12 . This means that both buyer and seller will “shade” their valuation by a

factor of 2/3 and there is a range of values of vs and vb for which no trade ever

takes place.

Further work: When does trade take place? When is it optimal for trade to

take place? Is the above mechanism efficient? Can you think of a more

efficient mechanism for bilateral trade?

15 Repeated games

In this part, we will consider contexts in which the same game is played a

repeated number of times, and we will show that repeating a game over several

time periods allows players to sustain equilibria that would not be equilibria of

the one-stage game and may improve per-period outcomes for both of them.

In this part, we consider settings in which all players know when the game will

end.

15.1.1 Example

2

L C R

T 3, 1 0, 0 5, 0

1 M 2, 1 1, 2 3, 1

B 1, 2 0, 1 4, 4

a) Suppose the game is played only once. Find all its Nash Equilibria.

Answer: This game has two PSNEs, (T, L) and (M, C). It also has mixed

strategy equilibria as follows:

15 REPEATED GAMES 56

1 − c − l respectively. Then expected payoffs for 1 can written as follows:

u1 (T ) = 3l + 5(1 − c − l)

u1 (M ) = 2l + c + 3(1 − c − l)

u1 (B) = l + 4(1 − c − l)

and one can see that u1 (T ) > u1 (B) so player 1 never plays B if player 2

plays a mixed strategy.

Consider now player 1 who mixes between T and M with probabilities t

and m respectively. Then expected payoffs for 2 can be written as follows:

u2 (L) = t + m

u1 (C) = 2m

u1 (R) = m

is never played in a mixed strategy equilibrium. One is thus left with the

following game to examine in order to find MSNEs of this game:

1

L C

2 T 3, 1 0, 0

M 2, 1 1, 2

3l = 2l + (1 − l)

1 = 2(1 − t)

1 1 1

so the unique MSNE of this game is such that l = 2, c= 2, t= 2 and

m = 12 .

b) Suppose now the game is repeated twice, that the players can observe the

outcome of the first stage before the second stage begins and the per-

period discount factor is δ. Show that the payoff (4, 4) can be achieved in

the first stage in a pure-strategy subgame-perfect Nash equilibrium, and

describe the conditional strategies which achieve this.

Answer: Consider 2’s conditional strategy such that if 1 plays B in the first

stage, 2 plays L in the second stage, and if 1 plays T in the first stage, 2

plays C in the second stage.

This strategy can be explained as such: 2 plays R in the first stage, hoping

that 1 will play B. If 1 plays B indeed, then he is rewarded by 2 playing

L in the second stage, when 1 will play T and get 3. If 1 plays T instead

(best response to R), then he is punished by 2 playing C in the second

stage, when 1 will play M and get only 1.

1 will conform to 2’s expectations whenever the discounted payoff to 1

under this equilibrium, 4 + 3δ is more than the deviation payoff of 5 + δ.

This is true for δ ≥ 1/2.

15 REPEATED GAMES 57

Now, suppose that as under this equilibrium, 1 plays B in the first stage.

Then the best response for 2 is to play R. There is thus no incentive under

this equilibrium for 2 to deviate from the first stage action.

Therefore, there is a PSNE of the two-stage game such that:

• 1 plays B and 2 plays R in the first stage,

• If 1 played B in the first stage, then 2 plays L in the second stage

(so 1 plays T in the second stage).

• If 1 played T in the first stage then 2 plays C in the second stage (so

1 plays M in the second stage).

This equilibrium is supported for any δ ≥ 1/2. One can check there is no

incentive for players to deviate from its prescribed course of action.

15.1.2 Definition

over T periods is defined by

1. Strategic form game G = N, {Si }i∈N , {ui }i∈N .

2. T , the number of periods over which the stage game G is played.

3. st the strategy profile played in period t.

t=T

4. H = ∪ S t , the set of possible histories in the game, with S 0 = ∅ the

t=0

initial history.

t=T

P t−1

5. Ui = δ ui (st ), the set of payoffs of the game, with st the strat-

t=1

egy profile played in period t, and δ the per-period discount factor.

Γ = G, H, {Ui }i∈N defines the finitely repeated game based on G

game with perfect information where the player function is the same at

every stage of the game and includes all players (P (h) = N for any h in

H).

• N = {1, 2},

• S1 = {T, M, B} , S2 = {L, C, R} , while

• {ui }i∈N are as shown in the normal form representation of the game

(for example, u2 (T, L) = 1).

15 REPEATED GAMES 58

t=T

• H = ∪ S t , the set of possible histories in the game, has got elements

t=0

such as for example {(T, L), (M, C)} .

t=T

δ t−1 ui (st ) with st the strategy profile played in period t. For

P

• Ui =

t=1

example, in the example above, U1 = u1 (T, L) + δu1 (M, C) = 3 + δ.

15.1.3 Exercise

W X Y Z

A 2, 2 x, 0 −1, 0 0, 0

B 0, x 4, 4 −1, 0 0, 0

C 0, 0 0, 0 0, 2 0, 0

D 0, −1 0, −1 −1, −1 2, 0

where x > 4.

a) Assume that the game is played only once. Find all its pure strategy Nash

equilibria.

Answer: There are three PSNEs: (A, W ), (C, Y ) and (D, Z).

b) Assume now that the game is repeated N times, that the players can observe

past outcomes before the current stage begins and that the discount factor

is δ.

achieved in the first stage in a pure-strategy subgame-perfect Nash

equilibrium. Describe the strategies which achieve this.

Answer: (B, X) can be supported in the first stage if 1 does not deviate to

A to get x and 2 does not deviate to W to get x. The best punishment if 1

deviates is for 2 to play Y in the second stage, and the best punishment if 2

deviates is for 1 to play D in the second stage. The reward if both conform

is to play A and W respectively in the second stage. This equilibrium is

supported if both 1 and 2 conform, which happens if 4 + 2δ ≥ x, which is

rewritten as δ ≥ x−4 2 . Since δ ≤ 1, this is possible only subject to x ≤ 6.

ii) Suppose that x = 9. What is the smallest number of N such that the

payoff (4, 4) can be achieved in the first stage in a pure-strategy

subgame-perfect Nash equilibrium. Describe the strategies which

achieve this.

Answer: Since x = 9 > 6 then (B, X) cannot be supported in the first

stage of a two stage game (see (ii)). Now, the minimum N that supports

(B, X) in the first stage must be such that 4+2δ+2δ 2 +2δ 3 +. . . 2δ N −1 ≥ 9.

If this is supported for δ = 1, then this is sufficient for our proof. Now,

15 REPEATED GAMES 59

in the first period. If (B, X) is played in the first period, then this will be

followed by (A, W ) in all the following periods. If 1 deviates to A in the

first period, then (C, Y ) will be played in all the following periods, while

if 2 deviates, then (D, Z) will be played in all the following periods.

Consider in this part situations in which the time horizon for both players is

infinite, so they do not know when the game will end. In that case, solving by

backward induction does not work, since there is no final period to start the

backward induction from.

15.2.1 Example

C D

C 5, 5 0, 7

D 7, 0 1, 1

The one-stage PSNE of this game is (D, D). This is the unique NE of this game.

Consider now a finitely repeated game of this stage game, with end period T.

One can check that by backward induction, the only SPNE of this stage game

is to play (D, D) in all periods. Does this translate into the infinitely repeated

version of the prisoners’ dilemma game? One will see that under some conditions

on the discount factor, the outcome (C, C) can be obtained in every stages of

the repeated game if T is infinite.

Definition: Convex Hull: Define the Convex Hull as the range of payoffs that

can be attained by playing pure-strategy combinations of the stage game,

then the convex hull of the game can be represented as follows:

15 REPEATED GAMES 60

u2

1 5 7 u1

For example, if players play (C, D) with probability 21 , (D, C) with prob-

ability 14 and (D, D) with probability 14 , then 1 will get expected payoff

of 2 and 2 will get expected payoff of 15

4 = 3.75. This is the point that is

represented inside the hull.

Nash-Threats Folk Theorem: Every payoffs within the convex hull that are

higher than the Nash equilibrium payoff profile for both players can be

achieved as a SPNE of the infinitely repeated game subject to δ being

high enough.

Remark: Note the proviso that δ be high enough. This means that some

payoffs that may be achieved for some δ may not be achievable for some

lower δ, and conversely. There are thus two ways to apply the theorem:

either see what is the maximum payoff that may be achieved for a given

δ, or see what level of δ is needed to attain a given payoff.

Example: The payoffs that are attainable can be represented as follows in the

PD studied above:

15 REPEATED GAMES 61

u2

1 5 7 u1

The convex hull minus its hashed areas is the set of attainable payoffs.

We will see below how such payoffs may be achieved.

Example: In the PD under study, let us for example study how the payoff

(5, 5) can be attained.

Consider the following ‘tit-for-tat’ conditional stage game strategy: Play

C in the first period. In the tth period, and if the other player played C

in all previous periods, then play C as well. If the other player played D

in any of the previous periods, then play D.

Playing C will then be optimal only if in all previous periods both players

always played C, and if playing C in this period obtains higher payoff than

playing D.

Let us see therefore what is the payoff to playing D: If one plays D in this

period, then one makes 7 this period, and 1 in all subsequent periods. If

one plays C in this period, then one will also play C in the next period

(the incentives are the same next period as in this period), so the expected

payoff of playing C this period is 5 forever. Therefore, playing C this

period is optimal s.t.

5 δ

5 + 5δ + 5δ 2 + 5δ 3 + .... = ≥7+

1−δ 1−δ

which translates in

1

δ≥

3

15 REPEATED GAMES 62

As claimed in the theorem, one can attain payoff (5, 5), which is inside the

convex hull and more than what is obtained in the NE of the stage game

for both agents, s.t. δ high enough, in this case, δ ≥ 13 .

that can be attained? The theorem tells us that a payoff within the non-

shaded area in the convex hull on the graph above can be attained, but

how, and what payoff?

The ‘how’ is rather complicated, and involves assuming that players have

access to publicly observable randomizing devices so that for example,

they know when to play C and when to play D. This public randomizing

device would draw their strategy at random according to pre-determined

probabilities. It would be possible for either player to check the other

player played what they were supposed to play. We can prove that the

greatest payoff that may be achieved subject to a given δ is such that

both players play symmetric strategies, i.e. both play C with the same

probability. The probability p with which both 1 and 2 are asked to play C

determines their payoff a. Indeed, I will have a = 5p2 +7p(1−p)+(1−p)2 .

a

1 will not deviate from playing C when told to do so s.t. 5p + δ 1−δ ≥

δ

7p + 1(1 − p) + 1−δ (the constraint for 2 is the same). I can thus conclude,

replacing a by its expression in p and rewriting the above equation, that

I will need that p be set such that a−1 1+p ≥

1−δ

δ . Note that, as stated

1

previously, for any δ ≥ 3 then p = 1 and full cooperation can be sustained.

For δ = 14 , no cooperation is sustainable (p = 0).

63

Part III

Incentive theory

Notations used in this lecture are drawn from Laffont J-J. and D. Martimort,

2002, The Theory of Incentives, Princeton University Press.

16 INTRODUCTION 64

16 Introduction

In this lecture, we consider situations with asymmetric information. A principal

is faced with an agent with whom he has to enter into a contractual relation-

ship. He does not know either the level of ability of the agent (adverse selection,

whereby inept agents are mixed with able ones) or he cannot control the level

of effort expended by the agent into the delivery of the contracted performance

(moral hazard then arises from the part of the agent, who may be tempted to

slack). Issues such as those give rise to incentive (or contract) theory, whose

aim is to examine how best to organize the principal-agent relationship in con-

texts of asymmetric information. Domains of application for contract theory

are outlined, and some typical agency problems are presented, as well as their

solution.

Asymmetric information is different from imperfect information. Instead of all

agents on the market holding the same set of (possibly imperfect) information

on every other agents in the market, some agents know more than others. In-

formation that can be found out at some cost is not asymmetric information.

Asymmetric information arises when agents are not motivated to reveal informa-

tion they hold, for example because that information provides them with some

advantage in a relationship. We will see how a principal can set up contractual

relationships such as to motivate the agent to reveal that information, at some

cost to the principal.

There are many different types of asymmetric information. Asymmetric infor-

mation on the type of the agent leads to adverse selection problems. In that

setting, for example, the fact you are healthy or not depends on outside factors

and not on your habits. Asymmetric information on the effort of the players

leads to moral hazard. In that case, for example, the fact you are healthy or

not depends on your habits. In the first case, you will want to screen agents,

in the second, you will want to monitor their habits or give them incentives to

behave ‘correctly’, i.e. in your own interest.

Some examples of situations with moral hazard or adverse selection follow:

Example 1: When buying a car, the buyer does not know whether the quality

of the car is good or bad. The seller/owner knows the quality of his car.

Absent any way to prove the quality of his car, the seller of a good quality

car will suffer a discount due to the presence of bad quality car sellers in

the market. This might be an explanation for why cars lose so much value

even right after being bought new.

Example 2: When selling car insurance, the insurer does not know whether

the insured is careful or not in driving. The insured knows it, and those

who drive badly will be more motivated to buy insurance than others.

In the limit, the insurer should never sell insurance because only if it

underestimated the risk of the insured would it sell. The same type of

problem arises in borrowing and in health insurance, as well as in many

used goods markets.

Example 3: In education, is a bad grade due to low effort by the student

(moral hazard on the part of the student), to his/her intellectual limita-

17 READINGS 65

tions (type of the student) or to bad teaching (which can be either due to

low effort on the part of the teacher or to his/her low ability)?

Example 4: In the current credit crunch, is the poor performance of banks

due to bad management of the banks, opportunistic behavior on the part

of bankers, bad macroeconomic policies on the part of the governement,

or poor regulatory oversight? The following Dilbert character seems to go

for the later option:

It is often difficult in each of those example to agree on who has the most infor-

mation (e.g. health). There are a variety of contracts that are signed between

agents in those settings, which are meant to alleviate asymmetric information.

For example, bonus/malus systems for car insurance are used in France: if you

have an accident, your insurance price increases, which deters drivers from driv-

ing carelessly. US insurance companies impose medical checkups to screen their

insurees. Companies offer return guarantees so as to prove their trust in the

quality of their goods.

17 Readings

• Kreps, Chs. 16-17

• Varian, Ch. 25

• Mas-Colell, Chs.13-14

• Laffont J-J. and D. Martimort, 2002, The Theory of Incentives; the Principal-

Agent Model, Princeton University Press.

• Salanié B., 2005, The Economics of Contracts, MIT Press.

17 READINGS 66

17.2 Articles22

• Akerlof G.A., 1970, The market for “lemons”: Quality uncertainty and the

market mechanism, The Quarterly Journal of Economics 84(3), 488-500.

• Grossman S.J. and O.D. Hart, 1983, An analysis of the Principal-Agent

problem, Econometrica 51(1), 7-45.

• Puertz R. and A. Snow, 1994, Evidence on Adverse Selection, Equilibrium

Signalling and Cross-Subsidization in the Insurance Market, Journal of

Political Economy 102(2), 236-257.

• Rees R., 1985, The Theory of Principal and Agent, Parts I and II in Hey,

J. & Lambert P. eds, Surveys in the Economics of Uncertainty, Blackwell;

also in Bulletin of Economic Research.

• Spence A.M., 1973, Job Market Signalling, The Quarterly Journal of Eco-

nomics 87(3), 355-374.

• Taylor C.R., 1999, Time-on-the-Market as a Signal of Quality, Review of

Economic Studies 66, 555-578.

The theory of games with incomplete information and the theory of asymmetric

information can be applied to a range of economic problems, as follows:

a) Regulation

And Regulation, The MIT Press. (This is a comprehensive book on the

theory of regulation.).

b) Auctions

• Klemperer P., 2000, Why Every Economist Should Learn Some Auction

Theory, Microeconomics 0004009, Economics Working Paper Archive at

WUSTL, available on the web.

• Fama E.F., 1980, Agency problems and the theory of the firm, Journal of

Political Economy 88(2), 288-307.

• Holmstrom B.R. and S.N. Kaplan, 2001, Corporate Governance and Merger

Activity in the U.S.: Making Sense of the 1980s and 1990s, MIT Dept. of

Economics Working Paper No. 01-11.

• Jensen M.C. and W.H. Meckling, 1976, Theory of the Firm: Managerial

Behavior, Agency Costs and Ownership Structure, Journal of Financial

Economics 3(4), 305-360.

22 Some articles in this list were contributed by previous teachers in MSc Economic Theory

1 at the UEA.

18 AGENCY AND ADVERSE SELECTION. 67

ment, Prentice Hall (This book is well written, accessible and motivating

and of great help to future managers).

d) Labour Economics

tests, Journal of Political Economy 91, 349-364.

• Hart O.D., 1983, Optimal Labour Contracts under Asymmetric Informa-

tion: An Introduction, Review of Economic Studies 50(1), 3-35.

• Holmstrom B., 1982, Moral Hazard in Teams, Bell Journal of Economics

13(2), 324-340.

• Laffont J-J. and M.S. Matoussi, 1995, Moral Hazard, Financial Con-

straints and Sharecropping in El Oulja, Review of Economic Studies 62,

381-399.

Press, second print 1998 (An excellent review for those interested).

• Rothschild M. and J.E. Stiglitz, 1976, Equilibrium in competitive insur-

ance markets: An essay on the economics of imperfect information, Quar-

terly Journal of Economics 90(4), 629-649.

• Stiglitz J.E. and A. Weiss, 1981, Credit Rationing in Markets with imper-

fect information, American Economic Review 71, 393-410.

f) Health Economics

• Arrow K.J., 1963, Uncertainty and the welfare economics of medical care,

American Economic Review 53(5), 941-973.

• Zechauser R., 1970, Medical Insurance: A case study of the trade off

between risk spreading and appropriate incentives, Journal of Economic

Theory 2, 10-12.

In this part, a typical model of adverse selection will be presented and the

timing of the contractual relationship will be outlined. The first best outcome

(which arises in contexts of perfect information) will be determined graphically

and compared with the second best outcome (with asymmetric information).

The role of incentive and participation constraints will be explained and the

maximization program of the principal will be determined as well as its graphical

solution. That solution will be compared with the first best.

18 AGENCY AND ADVERSE SELECTION. 68

produced by an agent at cost C(θ, q) = θq. θ is the type of the agent, known to

the agent and unknown to the principal, who expects θ = θH with probability

1 − v and θ = θL with probability v, with θH > θL .

The contracting variables C are (q, t), t being the (money) transfer from the

principal to the agent and q the quantity produced by the agent. Contract C

is enforceable under law, that is, a third party (a Court) can check a contract

was signed and can check whether t and q that were determined in the contract

were paid and/or produced or not.

Under contract C, the principal’s utility is then S(q)−t. Agent of type θ obtains

utility is t − θq.

The timing of the contractual relationship is as follows:

• At t = 2 the principal offers a (menu of) contract(s) C ( C1 , C2 , C3 , ...).

relationship (C = C0 = (0, 0) is always an option, as no one can be forced

into signing a contract).

• At t = 4 the contract is executed: the agent produces the quantity as

agreed in the contract and the principal pays her the transfer as agreed in

the contract.

Agent learns Principal offers Agent chooses q is produced

type θ contracts C contract t is paid

In a ‘first best’ world, the principal knows θ; he can therefore ask any quantity

from the agent provided the agent accepts. Knowing the type θ of the agent,

the program of the principal is:

maxS(q) − t (31)

q,t

s.t. t − θq ≥ 0

q,t

18 AGENCY AND ADVERSE SELECTION. 69

S 0 (q) − λθ = 0 (33)

−1 + λ = 0

t − θq = 0 (34)

so that

λ=1 (35)

0

S (q) = θ

t = θq (36)

Since λ > 0, the rent left to either type is 0 so type H will be offered contract

CH such that tH − θH qH = 0 while agent of type L will be offered contract CL

such that tL − θL qL = 0.

Given that λ = 1, the objective of the principal is then:

maxS(q) − θq (37)

q

When faced with an agent of type H, the principal will set qH s.t.

S 0 (qH ) = θH (38)

S 0 (qL ) = θL (39)

0 0

−1 −1

with qL = S (θL ) and qH = S (θH ).

The outcome of the maximization of the principal’s objective function can be

represented graphically as follows:

18 AGENCY AND ADVERSE SELECTION. 70

t tH- θ HqH=0

S(qH)-tH

tL- θ LqL=0

S(qL)-tL

tL FBL

tH FBH

q

qH qL

utility of the agent and of the principal, respectively, are maximized. Utility

P for the principal is represented by the iso-utility convex function such that

S(q) − t = P , which determines all combinations of q and t such that utility P is

achieved. Utility A for the agent is represented by iso-utility curve t − θq = A.

In the equilibrium of the first best, A is equal to 0 for both types of agents.

Utility for the principal is maximized at the point where the iso-utility convex

function of the principal is tangent with the iso-utility function of the agent, i.e.

the point such that t − θq = 0 and S(q) − t is maximized.

In the previous graph, one can observe that the utility of the good type is higher

if she chooses the first best contract of the bad type rather than the contract

that was destined for her. Therefore, the first best is not implementable if the

type of the agent is not known to the principal. Under the second best, each

type of agent must choose willingly the contract that is destined for him/her,

18 AGENCY AND ADVERSE SELECTION. 71

tH − θH qH ≥ tL − θH qL (ICH )

tL − θL qL ≥ tH − θL qH (ICL )

Under those conditions, type H chooses contract CH and type L chooses con-

tract CL .

tH − θH q H ≥ 0 (IRH )

tL − θL qL ≥ 0 (IRL )

if we want both types of agents to participate. Those are the individual ratio-

nality (‘IR’) constraints.

θH − θL > 0.

tL ,qL ,tH ,qH

s.t.

UH ≥ UL − ∆qL (41)

UL ≥ UH + ∆qH (42)

UH ≥ 0 (43)

UL ≥ 0 (44)

• First, note that adding up the participation constraint for both types, one

obtains:

UH + UL ≥ UL + UH + ∆(qH − qL ) (45)

which means that

qH ≤ qL (46)

The high type will produce less than the low type in the second best

outcome.

18 AGENCY AND ADVERSE SELECTION. 72

and the incentive constraint for type L is verified, then the participation

constraint for type L is verified. The participation constraint for type L

is thus superfluous.

• Third, suppose now the incentive constraint for type L is binding. Then

UL = UH + ∆qH , which, translated into the incentive constraint for type

L, means I must have UH ≥ UH + ∆(qH − qL ), which is always the case as

qH ≤ qL . Therefore, the incentive constraint for type L is also superfluous.

From this, I can conclude that the participation constraint for the high type

will be binding (UH = 0) while the incentive constraint for the low type will be

binding (UL = ∆qH ). This ensures all constraints are verified while minimizing

the rent given out to each type. Taking the two above inequalities and replacing

tH and tL by their expression in terms of qL and qH , the program of the principal

can thus be simplified into:

qH ,qL

S 0 (qL ) = θL (48)

and

v

S 0 (qH ) = θH + ∆ (49)

1−v

Note how the result is distorted compared to the first best maximization pro-

gram: There is no distortion for the good type who will produce the same as in

the first best, while the bad type will produce less than in the first best.

The outcome of the maximization of the principal’s objective function can be

represented graphically as follows:

19 AGENCY AND MORAL HAZARD 73

t tH-θHqH=0

tL-θLqL>0

S(qH)-tH

S(qL)-tL

tL SBL

tH SBH

q

qH qL

Interpretation of the graph: No rent is left to the bad type, who produces

less than the optimal level. The good type earns as much as what she would

earn if she chose the contract for the bad type, and she produces the first best

(optimal) level. The principal trades off between lowering the rent extracted by

the good type and still producing close to the efficient level. By lowering the

quantity asked from the bad type, the principal can decrease the rent that the

good type must be given. This is how we go from the transfers and quantities

that would be necessary to attain the first best, to the transfers and quantities

the principal asks for at the second best optimum.

In situations with moral hazard,

failure depend on effort which depends on the contract that is signed).

• The result from the contract is a noisy signal for effort, i.e. it is not possible

to know what effort was exerted from the result that was obtained.

• Transfers from the principal to the agent are constrained by the fact the

agent may have limited liability for its actions (i.e. there are limits on

what punishment you can impose in case of failure).

19 AGENCY AND MORAL HAZARD 74

the principal, and in that task the agent may exert effort e either 0 (no effort)

or 1 (effort). The cost of effort to the agent is ψ(0) = 0 for no effort and

ψ(1) = ψ > 0 for effort. The utility to the agent from getting transfer t and

exerting effort e is

U (t, e) = u(t) − ψ(e) (50)

with qH > qL , and the situation is such that

Pr(q = qH |e = 0) = p0 (51)

Pr(q = qH |e = 1) = p1 (52)

Success probabilities depends on effort and I will assume p1 > p0 so that a higher

effort brings about a higher probability of success – otherwise, why encourage

effort?.

The principal’s utility function is

with S(q) convex. The principal offers contracts such that the transfer will be

conditioned on the result:

is/are rejected, the game ends.

• At time t = 3, the agent decides whether to exert effort or not.

• At time t = 4, the outcome is realized

• At time t = 5 transfers are made according to the contract and the real-

ization of the outcome.

Principal offers Agents accepts Agent exerts Outcome is Payment

tH , t L or rejects effort or not realized is made

19 AGENCY AND MORAL HAZARD 75

fort)

Suppose the principal can observe effort and thus condition payment not on the

outcome but on whether effeot was exerted or not.

His objective function is:

qH ,tH ,qL ,tL

if he wants the agent to exert effort (the constraint ensures the agent’s ex-post

anticipated utility from accepting the contract and exerting effort is positive),

and

qH ,tH ,qL ,tL

if instead the principal wants the agent to exert no effort (this can happen if

compensating for a high effort is too expensive).

If effort is desired, and denoting λ the Lagrange multiplier, the Lagrangian can

be written:

L = p1 (S(qH ) − tH ) + (1 − p1 )(S(qL ) − tL )

(58)

+λ[p1 u(tH ) + (1 − p1 )u(tL )]

0

−(1 − p1 ) + λ(1 − p1 )u (t∗L ) =0 (60)

This means that t∗H = t∗L ≡ t∗1 and t∗1 is set such that the agent gets 0 utility so

u(t∗1 ) = ψ.

The same result is obtained if the principal chooses not to induce effort, but t∗0

is set such that u(t∗0 ) = 0.

The principal will chooses to induce effort only if:

19 AGENCY AND MORAL HAZARD 76

Suppose the agent has infinite wealth, so any tL (punishment in case of bad

outcome) can be asked for. Suppose also the principal cannot observe the agent’s

effort or equivalently, that he cannot prove effort was low. Then the agent must

be induced to choose the desired level of effort, so incentive constraints must be

verified as well as participation constraints. Suppose high effort is desired. The

agent’s incentive constraint is then such that

Solving the principal’s maximisation program, the two constraints (Incentive

and Participation) will be binding. This gives a system of two equalities with

two unknown,

p1 tH + (1 − p1 )tL − ψ = 0 (66)

p0 ψ

t∗L = − (67)

p1 − p0

(1 − p0 )ψ

t∗H = (68)

p1 − p0

So the expected payment for the principal is

This is the same as would be paid if effort was observable, i.e. the principal

merely compensates the agent for her effort. One would also see that the same

expected payment (0) as in observable effort would be made in case no effort

was needed. This means that even with asymmetric information on effort, the

first best level of effort is implemented.

Note however this holds only when the agent is risk neutral and has infinite

wealth. Let us now see what happens if there is a liability constraint or if the

agent is risk averse.

Suppose the limited liability constraint is L, so the agent cannot lose more than

L. One will then have to impose limited liability constraints such that tH > −L

and tL > −L. L can be interpreted as the initial wealth of the agent.

20 EXTENSIONS 77

If − p1p−p

0ψ

0

> −L, then the problem is unchanged (the first best outcome can be

implemented).

If − p1p−p

0ψ

0

< −L, then the outcome without liability constraint cannot be

achieved, and one obtains that

t∗L = −L (70)

ψ

t∗H = −L + (71)

p1 − p0

ψ

As in the previous program, t∗H = t∗L + p1 −p 0

, but t∗L is now raised to −L. The

agent will get a limited liability rent, as her expected utility is now

ψ

= p1 −L−ψ (73)

p1 − p0

ψ ψ

= (p1 − p0 ) − L − ψ + p0 (74)

p1 − p0 p1 − p0

ψ

= −L + p0 >0 (75)

p1 − p0

ψ

= p1 (S(qH ) − ) + (1 − p1 )S(qL ) − L (77)

p1 − p0

which is less than the first best. Because of limited liability constraints in moral

hazards problems, the principal may not want to induce effort even when this

would be optimal in the first best with perfect observation of effort.

We assumed up to now that u(t) = t, that is, the agent was risk neutral. If the

agent is risk averse, then u(t) is strictly concave. Then the principal will want

to limit the difference between tL and tH because this then makes the contract

less risky, and thus more acceptable to the agent. Risk aversion thus induces a

distortion from the first best, that is, there are situations where the principal

will not want to induce effort (it is too costly to do so) even though this would

be optimal under full information, and would occur under risk neutrality.

20 Extensions

This part examines some hidden assumptions made in the previous parts. In

particular, we assumed that:

20 EXTENSIONS 78

1. Agents are able to commit to fulfilling the terms of the contract they sign.

What happens then if agents cannot commit not to renegotiate a contract

once the type of the agent (adverse selection) or the outcome of his effort

(moral hazard) is known?

2. A court of law could be called on to enforce the terms of the signed con-

tracts. What happens then if the principal may renege on the contract?

We will therefore look at those issues and see how standard theory applies

in those alternative settings when the principal cannot commit. We will also

examine other possible extensions, such as what happens if the principal is

better informed than the agent, rather than the opposite.

The following model with repeated adverse selection illustrates what happens

if the principal cannot commit not to renegotiate after a contract is accepted

or after an action is taken by the agent (chapter 9, section 3, Laffont and

Martimort, 2002). Consider thus the standard adverse selection model, and

assume it is repeated twice with the same agent and the same principal both

period.

V = S(q1 ) − t1 + δ(S(q2 ) − t2 ) (78)

is the total utility of the principal over the two periods (the second period payoffs

are discounted by δ).

U = t1 − θq1 + δ(t2 − θq2 ) (79)

is the total utility of the agent over the two periods, with the second period

payoff discounted by δ.

Assume the principal can commit not to renegotiate. Then the optimal contract

is to repeat the optimal one-period contract in the two period. However, this

SB ∗

then calls for the bad type to produce qH < qH . The principal would then

gain from renegotiating after the first period, once agents revealed their type

through their choice of contract, so as to get the bad type, H, to produce the

∗

first best level qH . However, if the principal did this, the good type would then

have to be compensated for not taking the contract of the bad type. Therefore,

if the principal cannot commit not to renogotiate type H ’s contract, then she

must leave a higher surplus to the good type than if it could commit.

Denote qτ t and tτ t the quantity and transfers asked from and paid to type

τ = H, L in period t = 1, 2. For good type (L) not to choose the contract

designed for the bad type if the bad type’s contract is due to be renegotiated in

∗

the second period so the bad type would produce qH , one must have:

∗ ∗ ∗

tL1 − θL qL + δ(tL2 − θL qL ) ≥ tH − θL qH1 + δ(tH − θL qH )

∗

UL ≥ UH + ∆(qH1 + δqH )

20 EXTENSIONS 79

with UL (UH ) the intertemporal utility of the good (bad) type, the right hand

side of the equation the utility of the good type if she chose the contract designed

for the bad type, and as usual, ∆ = θH − θL .

This is to be compared with the constraint with full commitment (if the principal

can commit not to renegotiate)

UL ≥ UH + ∆(qH1 + δqH2 )

∗

Since qH > qH2 the first constraint is more stringent than the second, which

means the principal loses from his inability to not to renegotiate. There may

be a loss of efficiency from that situation if δ is high. Indeed, the principal may

then prefer not to know the types of the agents in the first period (and thus

offer a pooling contract in the first period, whereby both types choose the same

contract), so as not to learn the type of the agent in the first period, and thus

not have any basis for a renegotiation in the second period, when the optimal

one period contract will be offered.

In the following, we explore what happens in the more serious problem where the

principal cannot commit not to renege on a contract. This may be for example

because there is no court of law that can force him not to do so. In a moral

hazard setting, this leads to a hold up problem (chapter 9, section 4, Laffont

and Martimort, 2002). Take indeed the standard moral hazard problem. The

principal would like to commit to pay more when the result is good. However,

after the effort is made and a high result is achieved, the principal’s self interest

is to renege on the contract and not compensate the agent for his effort. This

problem can happen in a wide variety of settings even when an efficient court of

law is present. For example, this can happen if no formal contract was signed

beforehand or if the result to be achieved was ill-defined. This hold-up problem is

particularly prevalent for the State, especially if the judiciary is not independent

of the executive. The cost of forcing the State’s to abide by its commitments

can indeed be prohibitive. It is for example too easy for the State to change its

taxation policy to overturn previous pro-business commitments, or to withdraw

financial incentives to firms that adopted socially or environmentally-responsible

practices once firms have changed their way of doing business.

In those cases, the result of the principal’s inability to commit not to renege

on a contract is even worse for the principal than in the case where there was

inability to commit not to renegotiate: no contract can be signed with the agent

and agents exert only the lowest effort level. The problem may be mitigated if

rather than signing contract, the agent and the principal bargain after the state

of nature is revealed. It may also be mitigated if the agent knows what is the

state of nature before signing the contract and exerting effort.

20 EXTENSIONS 80

20.3.1 Informed principal

Situations with an informed principal are those in which it is the type of the

principal that matters, and this type is not known to either the principal or the

agent before contracting (we will see later about the case where the principal

knows the type). The situation can be modeled as follows:

• U = u(t − θq), is the payoff of the agent. As can be seen, it depends on

the type of the principal.

Principal offers Agent chooses Principal learns q is produced

contracts C contract type θ t is paid

Note that to the difference of the standard incentive problem, the principal

must now have an incentive to reveal his own type truthfully. Once the principal

reveals his own type, the agent produces q and receive t as agreed in the contract

for when the principal is of that specific type. The situation is reversed compared

to the standard, informed agent setting: if the principal is of a bad type, then

the first best level of production will be asked, while if the principal is of a good

type, then there will be more production than first best optimal (chapter 9,

subsection 1, Laffont and Martimort, 2002). Note however that if the agent is

risk neutral, then the first best levels of production can be achieved, while in

the standard case, even if the agent was risk neutral, there was still a distortion.

In the above, we focused on a case where, at the time of offering the contract,

the principal did not know his own type. Myerson (1983)23 examines the more

complicated case where the principal knows his own type before offering the

contract. This type of situation can happen for example in franchise contracts.

A potential MacDonald franchisee may know less about her business than the

franchiser (MacDonald), who indeed would know more about local market con-

ditions or the current success of its offering. This type of situation may also

apply to a job applicant when negotiating an employment contract with a firm.

The agent will then wants to learn the type of the firm (close to bankruptcy or

not, for example) from the type of contract the firm offers. Complex issues of

signaling, as in the ‘lemons problem’ for example, arise, as the principal’s choice

of contract may reveal his own type to the agent. There are two main cases

from the point of view of the theory: ones when the type of the principal only

enters the principal’s utility function,24 and ones when the type of the principal

also enters the agent’s utility function.25 That later case is the one we looked

at above.

23 Myerson R.B., 1983, Mechanism Design by an Informed Principal, Econometrica 51(6),

1767-1797.

24 Maskin E. and J. Tirole, 1990, The Principal-Agent Relationship with an Informed Prin-

25 Maskin E. and J. Tirole, 1992, The Principal-Agent Relationship with an Informed Prin-

21 THE REVELATION PRINCIPLE. 81

Seldom does one have pure adverse selection or pure moral hazard. In most

situations, there is either adverse selection followed by moral hazard or moral

hazard followed by adverse selection (chapter 7, Laffont and Martimort, 2002).

Insurance is a typical example: the effort by the agent in avoiding accidents

may depends on its type, her tendency for recklessness for example. In this

case, adverse selection (type) precedes moral hazard (effort).

The adverse selection problem will be exarcebated by moral hazard (compound-

ing factor), as the bad, reckless type will do less effort to avoid accidents, thus

increasing the discrepancy in terms of probability of a bad outcome between

good and bad type. In the opposite case, where moral hazard is followed by ad-

verse selection (for example, when effort to get an education also changes one’s

type in a way that is favorable), then the moral hazard problem may be weak-

ened by adverse selection, as agents may exert more effort in order to acquire a

good type, and not only just to increase the probability of a good outcome.

Contracts are generally simpler than what theory would call for. Laffont and

Matoussi (1995)26 examine real world sharecropping contracts and assess the

loss of efficiency due to their relative simplicity. Are simple contracts signed

because there is too much of a cost to write complicated contracts? What then

does that cost consist in? Tirole (2009)27 is a recent contribution exploring this

topic.

It was assumed that the agent knew its own type (adverse selection), or that it

was able to choose its own level of effort (moral hazard). What happens now if

the participants can choose what information they have access to and at what

stage? This calls for a design of the contracts informational structure as well as

of the contracts. It then becomes necessary to motivate information acquisition

by the agent and/or the principal.

In adverse selection problems we assumed up to now that the principal would

offer one contract (or less) for each different type of agent. For example, with

two types, the principal would offer either:

• Two different contract that would be such that each type would choose a

different one (the contracts are ‘separating’ both types), or it would offer

only

26 Laffont J-J. and M.S. Matoussi, 1995, Moral Hazard, Financial Constraints and Share-

27 Tirole J., 2009, Cognition and Incomplete Contracts, American Economic Review 99(1),

265-294.

21 THE REVELATION PRINCIPLE. 82

• One contract that both types would choose (the contract is ‘pooling’ both

types).

• One contract such that only one type would accept a contract (the contract

is excluding one type). The other type chose the ‘no contract’ option

C = (0, 0) that gave it its reservation utility.

In the case of separating contracts, as each type chose a different contract, the

principal gained information on the agents’ types from their choice of contract.

This section re-examines the very simple setting introduced in the lecture on

adverse selection to examine whether it could ever be beneficial for the principal

to offer a third or more contracts, offer more options than simply quantity and

transfer (q, t) contracts or asking the agent for more information than simply

their choice of contract.

Stating the above in another way, we restricted ourselves up to now to Truthful

Direct Revelation Mechanisms (‘DRM’) whereby:

on θ, the type announced by the agent. This is the mechanism.

2. Agents were only required to directly reveal a type, so agent of type θ who

would announce θ0 would produces q(θ0 ) and get t(θ0 ) . Her utility was

then U (θ, θ0 ) = t(θ0 ) − θq(θ0 ).

3. The principal designed her contracts such that agent of type θ would

announce her type was θ (truthful ) and got contract C(θ). This was

guaranteed by the incentive constraint:

for any θ0 6= θ

In the case where there was for example ‘bunching’ or pooling of type θ

and θ0 , there was no loss for the agent in announcing θ, rather than for

example θ0 , as C(θ) = C(θ0 ). This means there was at least weak incentive

to reveal one’s type.

This part will prove that the above type of solution, in terms of truthful DRM

is general, that is, it is never beneficial to offer more contracts than there are

types (principle of economy), there is no point in asking the agent for more than

to just announce a type (or for more than just choosing a contract), and there is

no loss in generality in looking only at DRM where agents announce their type

truthfully.

Obviously, if the only communication between the agent and the principal is to

tell one’s type (or another type), then there can be only so many contracts as

there are types.28 In order to test the principle of economy, let’s thus consider

28 Unless, to the announcement of a type, the principal offers a contract chosen at random

between a number of contracts. But then, among those contracts, there must be one that

gives higher utility to the principal than another, given the announced type, so the principal

would be better off not choosing randomly.

21 THE REVELATION PRINCIPLE. 83

Those messages can be the choice of contract, the type of the agent, or any

other information. The mechanism is then of the form C : M → A, with A the

set of allocations (q, t). This means that to any message m in M , the function

C associates an allocation (contract) s.t. C(m) = (q(m), t(m)). Agent θ will

then rationally choose to communicate the message m∗ (θ) that maximizes her

surplus, that is

m∗ (θ) = arg max{t(m0 ) − θq(m0 )}. (81)

m0

need for such complication in the mechanism, so (1) each agent can limit her

communication to announcing her type (rather than some more general message

m), and (2) there is also no need to consider mechanisms whereby an agent would

announce a type different from her own.

Denoting Θ the set of types, (1) is shown by saying that instead of having a

mechanism such that one type send her optimal message and the message maps

into an allocation A:

m∗ (θ) C(m)

Θ → M → A (82)

one can simply consider C 0 (θ) = C ◦ m∗ (θ) whereby the principal replaces C by

the equivalent menu of contract C 0 that maps types directly into allocations:

C 0 (θ)=C◦m∗ (θ)

Θ → A (83)

(2) is shown by saying we can limit ourselves to truthful direct revelation mech-

anism, whereby agent announce their own type. Indeed, replacing m0 by m∗ (θ0 )

in equation (81) one obtains:

m0

= arg max{t(m∗ (θ0 )) − θq(m∗ (θ0 ))] (85)

θ0

and since t(m∗ (θ)) = t0 (θ) under the menu of contract C 0 introduced above,

this means that in the same way as the agent’s message truthfully revealed the

type of the agent under contract C, the type announced under contract C 0 is

truthful.

Intuitively, suppose the contracts offering were such that an agent of type θ is

better off announcing she is of type θ0 6= θ (all other agents announce truthfully).

Then this contract offering is equivalent to another contract offering that gets

rid of the contract that is assigned to an agent that announces she is of type

θ. Under that new contract offering, agents θ and θ0 are pooled into the same

contract. This new contract offering is equivalent to a contract offering such

that whether you announce your type is θ or θ0 , you are assigned the same

contract. And in this case, agent of type θ may as well announce she is of type

θ. Therefore, by extension, any contract offering that is not a truthful DRM

has an equivalent that is a truthful DRM.

21 THE REVELATION PRINCIPLE. 84

can be used to great effect to simplify and solve apparently intractable problems.

We also consider limitations to the truthful DRM in voting problems, where

there are many agents and agents know not only their own type, but also know

others’ types.

We saw in the game theory part of this course that offering agents to play second

price auctions resulted in a DRM where each agent announced her type (i.e. her

valuation for the good). If she won (her valuation was the highest), she paid

the second highest announced valuation. First price auction however did not

result in a DRM: agents, rather than announcing their own valuation, would

announce half their own valuation (indeed, each agent would optimally bid half

her valuation for the good). However, first price auctions are implementable

through the following DRM: agents announce their own type (valuation for the

good), and pay half of that valuation if they win. More generally, the revelation

principle is useful in the context of auctions as it shows there is no point in

studying such and such specific mechanism (first price, second price, loser pays,

etc.), as one can limit oneself to simple, direct revelation mechanism, knowing

that they are the equivalent of the more complicated designs.

What we said up to now was that a principal who wished to make a decision

when faced with informed agents could limit himself to considering only mech-

anisms (ways to reach a decision) that are truthful DRMs. However, this does

not mean that such mechanisms will have any desirable properties, or even that

they will exist.

In order to illustrate this point, consider the following collective decision prob-

lem. There are three agents, a, b and c who have to collectively choose between

option, x, y, z or w. Suppose agents rank options according to the following

table:

Preferences

a xyzw

Agents b xyzw

c zyxw

worst and w as worst.

1) Suppose a principal knows the agents’ preferences and wishes to design a

mechanism such that the chosen option will maximize collective welfare. In this

case, this can be interpreted as minimizing the sum of the rank of the option

chosen across all agents. This is the Borda rule, named for Jean-Charles de

Borda, who devised the system in 1770 for elections to the French Academy of

21 THE REVELATION PRINCIPLE. 85

sums up to 12. The optimal decision would thus be x.

2) Suppose now however that the principal does not know the ranking of each

option by each agent. Suppose however that the agents know their own rankings,

and the rankings of others. One could then offer agents to choose according to

the Borda rule. Each agent should announce (secretly, that is, only to the

principal) a ranking of alternatives. The principal should choose the option

that minimizes the sum of the rank of the alternatives across all agents.

Suppose thus that under this mechanism, all agents announce their rankings of

options truthfully. Then as seen above, option x would be chosen. However,

agent c could change to announcing y z w x instead of z y

x w, thus lowering the overall rank of x and increasing the rank of y. The

announcements would then be as follows:

Preferences

a xyzw

Agents b xyzw

c yzwx

y now scores 5, while x scores 6, so that by this manipulation, c gets its second

preferred choice adopted, rather than its third choice.30 Anticipating this, a or

b could change their choice, leading to a cycle of changes away from the truth.

It is possible to generalize from this, and to say that one cannot implement a

truthful DRM in such a voting situation, except for the dictatorial rule (one

agent is given the role of a dictator and chooses for others. He then chooses his

most preferred option. This is a DRM). Obviously, this dictatorial rule is very

inefficient, since it fits the preference of only one person.

This is the subject of the Gibbard-Satterthwaite theorem (Chapter 23.C of the

Mas-Colell),31 which is similar (if not identical)32 to the very interesting Arrow’s

impossibility theorem (Chapter 21.C of the Mas-Colell).33

This means that there is no hope for a rule that would lead the principal to

choose the best option for all possible rankings of the options by all agents. This

is because any rule could be subject to manipulation under some circumstances;

some agents would prefer to lie about their type (ranking of alternatives).

29 de Borda J-C., 1781, Mémoires sur les élections au scrutin, Histoire de l’Académie Royale

30 To the credit of Borda, he knew his scheme was subject to manipulation, saying ‘My

31 Gibbard A., 1973, Manipulation of Voting Schemes: A General Result, Econometrica 41,

587-601.

Satterthwaite M.A., 1975, Strategy-Proofness and Arrow’s Conditions: Existence and Corre-

spondence Theorems for Voting Procedures and Social Welfare Functions, Journal of Economic

Theory 10, 187-217.

32 Reny P.J., 2000, Arrow’s Theorem and the Gibbard-Satterthwaite Theorem: A Unified

33 Arrow K.J., 1950, A Difficulty in the Concept of Social Welfare, Journal of Political

21 THE REVELATION PRINCIPLE. 86

While this is true if agents know each other’s type (ranking), there are truthful

DRM in Bayesian strategies if agents know their own type (ranking), but do

not know the type (ranking) of the other agents (Chapter 23.D of the Mas-

Colell). In that case there are mechanisms such that the agents announce their

type truthfully given their expectation (the average) of what the other agents’

types will be. Note that to the different of truthful DRM that we looked at

up to now, where the agent would play the same whathever his belief about

other agents’ types (truthful DRMs in dominant strategies), truthful DRM in

Bayesian strategies lead the agent to play differently depending on his belief

about other agents’ type. Truthful DRM in Bayesian strategies are thus less

robust than truthful DRMs in dominant strategies, as their outcome will depend

on agents’ beliefs, which may be arbitrary and out of step with reality.

87

Part IV

Exercises

22 CHOICE UNDER UNCERTAINTY 88

1. The figure shows indifference curves for a particular individual over lotteries

involving three possible outcomes, x1 , x2 , and x3 , such that x3 is strictly

preferred to x2 which in turn is strictly preferred to x1 .

Figure 17: Two lotteries and indifference curves in the Marschak-Machina tri-

angle

a) how their attitudes towards risk differs between point A and point

B in the figure and

b) which of the axioms of expected utility theory are broken by these

indifference curves.

• Lottery C pays £10 on numbers 1-4, and nothing on 5-6.

34 Exercises 1, 2 and 3 were contributed by previous teachers in MSc Economic Theory 1 at

the UEA.

22 CHOICE UNDER UNCERTAINTY 89

b) Are they consistent with prospect theory?

c) Are they consistent with regret theory?

• Lottery C pays £10 with probability 0.25 otherwise nothing.

• Lottery D pays £30 with probability 0.1, otherwise nothing.

• Lottery E pays £12 with probability 0.5, otherwise nothing.

• Lottery F pays £30 with probability 0.2, £12 with probability 0.6, other-

wise nothing.

Adelina’s preferences over lotteries satisfy the axioms of expected utility the-

ory. She prefers lottery C to D. What are her preferences over the following

pairs of lotteries?

a) C versus E

b) D versus E

c) E versus F

d) A versus B

Suppose the individual rejects a 50-50 chance of losing £10 and winning

£11 at an initial wealth level w.

a) Show that this implies that ln(w − 10) + ln(w + 11) ≤ 2ln(w).

b) Deduce from a) that w ≤ 110 (Reminder: ln(ab) = ln(a) + ln(b) and

cln(a) = ln(ac ))

22 CHOICE UNDER UNCERTAINTY 90

offered a 50-50 chance of losing £100 and winning £Y . What is the

minimum Y such that the individual accepts?

is offered a 50-50 chance of losing £100 and winning £Z. What is

the minimum Z such that the individual accepts?

e) Comment on the above results. Do you find the predictions from c)

and b) reasonable? Justify your answer.

with probability 1 − q. Suppose the initial wealth of the agent is w and

the agent is an expected utility maximizer.

a) Suppose the agent offers others to play the lottery and prices the

ticket for the lottery at pS . What is the minimum price that the

agent will set? Write down the condition on pS in terms of the

agent’s utility function u(.).

b) Suppose now the agent wishes to buy a ticket for the lottery. What

is the maximum price pB that the agent would be ready to pay for

that ticket?

c) Suppose that G = 10, L = 2, w = 10 and q = 0.5. Suppose also

that u(x) = ln(x). Compute the values of pS and of pB .

Consider a world with three possible states, S1 , S2 and S3 , that occur with

probabilities p1 , p2 and p3 respectively. An agent has the choice between

actions A, B and C, such that action A obtains outcome a if S1 occurs,

and d otherwise; action B obtains outcome e if S3 occurs, and b otherwise;

and action C obtains outcome c in any state of the world. Outcomes are

expressed in monetary terms and a > b > c > d > e.

Suppose the agent’s behaviour fit the axioms of Expected Utility Theory.

What are the expected utilities of A, B and C?

Explain your answer.

35 Loomes G. and C. Taylor, 1992, Non-transitive preferences over gains and losses, The

23 GAME THEORY 1 91

Suppose now the agent exhibits regret aversion. She associates regret r(x, y)

when she would have obtained outcome x in the lottery she chose while

she would have obtained outcome y in the lottery she did not choose.

When comparing two lotteries, L1 and L2 , she chooses L1 if the expected

regret of choosing L1 is positive. Assume that: r(x, y) is increasing in x;

r(x, y) = −r(y, x); and r(x, y) > r(x, z) + r(z, y) when x > z > y.

d) Write down the conditions under which the agent chooses: B over A;

C over B; A over C. Is this consistent with regret theory? Comment.

(Hint: sum up the three conditions you identified and simplify using

the properties of the regret function).

e) Write down the conditions under which the agent chooses: A over B;

B over C; C over A. Is this consistent with regret theory? Conclude.

36

23 Game Theory 1

1. In the following game, find all pure strategy Nash equilibria:

L R

L 6,6 0,8

R 0,3 2,2

W X Y Z

A 4,6 0,4 8,4 9,2

B 4,2 2,6 6,4 8,4

C 4,2 2,8 4,9 6,0

D 0,5 0,3 1,7 2,6

gies.

b) Apply iterative deletion of weakly dominated strategies to the game

that remains after completion of (i).

c) Identify all pure and mixed strategy Nash equilibria of the game.

36 Exercises 1 to 5 were contributed by previous teachers in MSc Economic Theory 1 at the

UEA.

23 GAME THEORY 1 92

U D

l r

l r

0 4 2 -1

0 1 2 -1

b) Find all Nash and subgame perfect Nash equilibria.

1 R

2

2

L M

2 2

l r

l r

3 0 1 0

1 0 0 1

23 GAME THEORY 1 93

b) Find all Nash and subgame perfect Nash equilibria.

1

R

2

4

L M

2 2

l r l r

m m

1 1 4 4 0 3

3 2 0 0 2 3

c) Comment on the plausibility of these.

Up 10,2 -10,1 0,0 -10,-10

Down -10,-10 10,-5 1,0 10,2

24 GAME THEORY 2 94

a) Suppose that players are able to agree on their move before playing.

What would they choose?

b) From now on, suppose players 1 and 2 are not able to communicate

and must choose their moves independently. Suppose you are player

2. Intuitively, what seems like a reasonable way to play when you

are not sure what player 1 will choose? Explain.

c) What are the pure strategy Nash equilibria of this game?

d) Consider mixed strategy equilibria such that player 1 mixes between

Up and Down with probability q and 1 − q respectively.

i) Show that player 2 will never play Left for any value of q.

ii) Show that if player 2 plays mixed strategies and plays Stop with

some positive probability, then it plays Right with zero probability.

iii) Show that player 2 will play Middle whenever q belongs to the

interval [1/6, 5/6].

e) Rationalize your choice of action in (a) taking into account (d)(iii).

24 Game Theory 2

U D

1

3 L R

0 3

0 1

a) Write the game in normal form and find all of its Nash equilibria.

24 GAME THEORY 2 95

the intuition for this equilibrium.

play R. The decision of player two whether to pre-commit or not is

know to player 1.

i) Draw the new game in extensive form, where in a first stage player

2 decides to pre-commit or not.

ii) Find the Subgame Perfect Nash Equilibrium of this new game for

every possible value of c > 0.

iii) When does player 2 pre-commit? Explain.

2. Consider a buyer and a seller. The seller has only one unit to sell and values

the good at £50 while the buyer values it at £60 if it is of a high type and

£55 if it is a low type. The seller does not know the type of the buyer,

but knows there is proportion h of high type. The buyer knows its own

type.

a) Suppose the seller asks for a price p for the good and p is a take it

or leave it offer. The buyer can only either accept or reject the price

offer. Find out the seller expected payoff as a function of p. What

price will the seller set?

Suppose now the game is played over two periods, and the buyer’s discount

factor is t < 1 (the buyer is impatient) while the seller’s discount factor is

1 (the seller is patient). Suppose also the seller can change prices between

periods: p1 is the price asked in period 1 while p2 is the price asked in

period 2. Both periods involve the same buyer and the same seller, and

prices p1 and p2 are announced in the first period. The seller can commit

to prices p1 and p2 .

determining the type of the buyer.

c) Suppose 60 > p1 > 55 > p2 > 50. What is the expected payoff for a

high type buyer from accepting in the first period? From waiting?

What about the low type?

d) Show there exist p1 and p2 such that high type buyers buy in the

first period and low type buyers buy in the second. What conditions

must p1 and p2 satisfy? What is the expected payoff of the seller

under that type of equilibrium?

e) Suppose the seller cannot commit on p2 in period 1. Is the separating

equilibrium of question d) still sustainable? Why or why not?

25 INCENTIVE THEORY 96

3. Consider two players, A and B, who are engaged in a public goods game

whereby the public good is produced and has value 1 to both if either

or both of the players contribute to it, while it has value 0 if neither

contributes. This means the good is produced and its value is available to

all even if only one player contributes to it. The costs of contributing are

ca and cb for players A and B respectively. The term ca can take one of

the two values: C > 1 with probability 1 − λ and c < 1 with probability

λ; cb independently follows the same distribution. We call a player whose

cost of contributing is C the inefficient type, while if that cost is c we call

that agent the efficient type. Each player knows his/her type, but not the

other player’s type. Players choose at the same time whether to contribute

and do not observe whether the other player contributes.

type c is called efficient.

b) Explain why it is important that contributions be made at the

same time and players cannot observe whether the other player con-

tributes.

c) Suppose λ = 0 which means both players are efficient. Draw the

game in normal form. What are the Nash equilibria of the game?

What if λ = 1?

d) Suppose now 1 > λ > 0. Draw the game in extensive form (Hint:

Represent the draw of players’ type as a move by nature).

e) Depending on the type of the agents, what is the Pareto optimal

outcome of this game?

f) Show that in a Bayesian equilibrium of this game, inefficient player

A never contributes, while efficient player A contributes with some

probability pa . Express pa and pb as a function of the parameters

of the game. Be careful to explain the notations you use and the

reasoning you are following.

g) Suppose now the game is repeated twice, and players observe each

other’s action after the first period and before playing the game

again. Denote pta (ca ) the probability that player A of type ca con-

tributes in period t. Show again that a player of a low type never

contributes. How does player B interpret a contribution by player

B in period 1?

25 Incentive theory37

1. Consider an agency relationship in which an agent can take two efforts,

either low or high. The agent’s effort might result in two results, 10 or

100 depending on the state of nature.

37 Exercises 1, 3 and 4 were contributed by previous teachers in MSc Economic Theory 1 at

the UEA.

25 INCENTIVE THEORY 97

is summarised by one of the following matrices. In these matrices, each

entry is defined as the probability that the result is the COLUMN result

given that the effort is the ROW effort. So, for instance, in CASE A, 0.3

is the probability of the result being 100, given effort is LOW.

CASE A

effort/result 10 100

LOW 0.7 0.3

HIGH 0.1 0.9

CASE B

effort/result 10 100

LOW 0.1 0.9

HIGH 0.9 0.1

CASE C

effort/result 10 100

LOW 0.6 0.4

HIGH 0 1

Each matrix represents a different situation but in only one of them is there a

real moral hazard problem. Discard two of these matrices and carefully argue

why there is no moral hazard problem in them.

Consider a one-time production project that requires an initial capital

input worth K = 100, and then returns revenue worth R = 240 if the

project is a success, or returns no revenue (0) if the project is not a success.

The project’s probability of success depends on the manager’s hidden ac-

tion. If the manager diligently applies good effort to managing the project,

then probability of success is pG = 1/2. On the other hand, if the manager

behaves badly and abuses her managerial authority in the project, then

the probability of success is pB = 0, but the manager gets hidden private

benefits worth B = 30 from such abuse of power.

The manager owns no personal assets and does not invest anything in the

project.

a) Show that the expected returns from the project (including possible

private benefits) are greater than the cost of its capital inputs only

if the manager chooses to be good.

38 Myerson R.B., 2008, Perspectives on Mechanism Design in Economic Theory, American

25 INCENTIVE THEORY 98

nothing in case of failure. Under what condition on w does the

manager undertake to be good?

c) Suppose the investor chooses w such that the manager undertakes

to be good. What is then the highest possible return of the project

to the investor? Does the investor invest? Comment.

type 1, and the remaining proportion (1 − p) is of type 2. The payoff

for a consumer of type i = 1, 2 from consuming quantity q of the good

is Ui = q[ti − q/2] − R(q), with i = 1, 2, t2 > t1 and R(q) denotes the

payment made to buy q. The consumers’ reservation payoff (or consumer

surplus) is zero.

The monopolist wishes to discriminate between customers; would he know

the customers ‘types’ he would extract the full consumer surplus from

them. However, he is unable to do so, and manages to price discriminate

only up to the second degree. This means he will offer two contracts

destined for each type, with contract i = 1, 2 to be chosen by type i

denoted as Ci = (qi , R(qi )). The monopolist has a constant marginal cost

c and no fixed cost so the firm’s type-specific profit is Si = R(qi ) − cqi .

Your task is to determine the monopolist’s optimal quantity-payment

schedule that accomplishes second-degree price discrimination. Follow the

questions in sequence.

can distinguish between the types of the consumers. Show your

answer graphically (on a quantity- payment plane) and explain your

findings.

b) Introduce asymmetric information and show how the first best scheme

becomes sub-optimal using a graph.

c) Set the monopolist’s profit maximization problem formally with ap-

propriate self-selection and participation constraints. Identify which

constraints will bind, and which will not.

d) What is the meaning of “no distortion at the top” and “no rent at

the bottom” in this context?

e) Finally, solve for the optimal quantity-payment schedule and show

it on the graph.

4. Imagine that a hospital is contracting out its cleaning service. The contracted

firm is to supply a level I of cleanliness, which is observable. There are

two types of firms, which differ in their costs of producing cleanliness. For

the first type, the cost of providing I level of cleanliness is I 2 , while for

the second one it is kI 2 , with k > 1.

25 INCENTIVE THEORY 99

made dependent on I. The contractor’s profit is thus S = P − I 2 , or

S = P − kI 2 depending on its type. Assume that its reservation profit is

10. The hospital’s utility is U = AI − P , where A is a positive parameter.

a) Which firm is the “top” (efficient) type, which one is the “bottom”

(inefficient) type?

b) Assume that the hospital can observe the type of the firm. Write the

maximisation problem for the hospital if it contracts with the top

type. Find the first order conditions and the relationship between

P and I at the solution.

c) Assume that the hospital cannot observe the type of the firm. Write

the maximisation problem for the hospital if it wants the two types

of firms themselves to select different types of contract. Assume that

1/2 of catering firms are of each type.

d) Identify the constraints that will bind in the solution, and solve for

the optimal (P, I) schedule.

e) Present your answer to d) graphically.

5. Consider an agent with utility function U = u(q) − c(e) where u(.) is defined

over output and c(e) is the cost of effort e. Effort e can be either 0 or

1 and determines the probability of success in producing output Q > 0.

The cost of effort is C > 0 if effort is 1 and 0 if effort is 0. p1 is the

probability with which the agent exerting effort e = 1 is successful and

produces output Q > 0, while p0 < p1 is the probability with which the

agent exerting effort e = 0 is successful and produces output Q > 0. If

the agent is not successful, then he produces no output.

Consider a paternalistic employer that seeks to maximize the welfare U

of the agent, and can pay wages such that the agent that is successful

receives wage W and the agent that is not successful receives w. Note

that wages can be negative in which case the agent pays the employer.

Note also that the employer cannot make losses in expected terms.

i) Write down the agent’s participation constraint (the condition

under which expected payoffs for the agent is positive, taking into

account wages from the employer) and the agent’s incentive compat-

ibility constraint (the condition under which the agent exerts effort).

ii) Write down the employer’s objective function and maximization

program.

iii) Write down the employer’s necessary and sufficient Kuhn and

Tucker optimality conditions with respect to W and w. How much

does the employer pay in case of success and in case of failure?

Comment.

25 INCENTIVE THEORY 100

i) Write down the employer’s objective function and maximization

program.

ii) Write down the employer’s first-order condition for optimality

with respect to W and w. Show that W 6= w. Comment.

6. A firm that faces productivity shocks has the following profit function: π =

xh − w with x = xH with probability q (favourable shock) and x = xL

with probability 1 − q (adverse shock), xH > xL > 0, h the number of

hours worked by the employee of the firm and w the wage paid to the

worker.

The worker’s utility is u(w) − f (h), with u(.) concave, f (.) convex, w the

wage paid by the firm to the employee and h the number of hours worked

by the employee. Suppose the firm chooses the terms of the worker’s

employment contract, and the worker’s reservation utility is 0. The con-

tract is signed before realization of the shock; the productivity shock is

observed after contracting. Contract variable can be made dependent on

the realization of the shock as its realization is publicly observable and

the contract can thus be enforced by the Courts.

a) Who is the principal? Who is the agent? What variables are the

firm and the worker going to contract upon?

c) Write the principal’s objective function. Write the agent’s individual

rationality (IR) constraint.

d) Find the first order condition for the maximization of the princi-

pal’s objective function subject to the agent’s IR constraint. (Re-

minder on constrained maximization: The solution(s) to a con-

strained maximization program, with f (x, y) the objective function

and u(x, y) > 0 the constraint, is obtained by taking the derivative

of that program’s Lagrangian f (x, y) − vu(x, y), w.r.t. x, y and v,

and equating it to 0, with v the Lagrange multiplier).

e) Prove that the agent will receive the same wage regardless of the

realization of the shock. When does the agent work more?

f) Suppose the shock is not publicly observable, is known only to the

firm and the firm can lie about the realization of the shock. Suppose

it offers the contract determined in e) above. Will the principal lie

about the realization of the shock? When? Why?

g) Write down the conditions under which the principal has no incentive

to lie. Using only those conditions, compare the wage and the hours

worked in case of an adverse and in case of a favourable shock.

101

Part V

Correction of exercises

26 CHOICE UNDER UNCERTAINTY 102

1.

a) The steeper the indifference line, the more risk averse the agent is.

See page 126, note 6 of Machina (1987).39 Intuitively, suppose for

example that the vertices of the triangle are £0, £50 and £100. If

the agent is risk neutral, then she values a lottery (1/2, 0, 1/2) the

same as (0, 1, 0) and the indifference line linking those two points

thus represents the preferences of a risk neutral agent. A risk averse

agent will prefer the second lottery to the first, so her indifference

line will be steeper. The opposite holds for an agent that is risk

loving.

b) The independence axiom requires that indifference lines be paral-

lel, but here they are not parallel, so the independence axiom is

not respected. Other axioms are not broken, as there is no cross-

ing of indifference curves within the triangle, indifference curves are

continuous and they span the whole area of the triangle.

2.

1 2 1

2 u(0). Multiplying by 3 both sides and adding 3 u(0) on both sides,

2 1 1 2

one finds 3 u(10) + 3 u(0) ≥ 3 u(24) + 3 u(0), which translates into C

preferred to D.

Therefore, the combination A preferred to B and D preferred to C

is not consistent with EUT.

b) For A preferred to B and D preferred to C to be consistent with

prospect theory we must have both:

1 1

π(1)u(10) ≥ π( )u(24) + π( )u(0) (86)

2 2

and (87)

2 1 1 2

π( )u(10) + π( )u(0) ≤ π( )u(24) + π( )u(0) (88)

3 3 3 3

Normalizing u(0) = 0 (this is allowed as the utility function is unique

up to a linear transformation) this can be rewritten into the following

condition:

π(1) u(24) π( 23 )

≥ ≥ (89)

π( 12 ) u(10) π( 13 )

I will show that this can occur under some specification of the

function π(.). Suppose π(.) is a continuously increasing function,

39 Machina M., 1987, Choice under Uncertainty: problems solved and unsolved, Journal of

26 CHOICE UNDER UNCERTAINTY 103

π( 31 ) = 13 , π(x) > x for any x < 13 and π(x) < x for any x > 31 (this

fits common assumption about the shape of the π(.) function and

corresponds to some empirical findings), and suppose u(x) = x for

any x (risk neutral agent). Note also that π(1) = 1. The condition

above can then be rewritten as

1 24 2

≥ ≥ 3π( ) (90)

π( 12 ) 10 3

2.4 1 1 10

3 from the fact π(x) < x for any x > 3 , while π( 2 ) ≤ 24 , while

more stringent than what we assumed, does not contradict PT.

c) Under regret theory, A preferred to B translates in

1 1

r(10, 0) + r(10, 24) ≥ 0 (91)

2 2

while D preferred to C translates in

1 2

r(24, 0) + r(0, 10) ≥ 0 (92)

3 3

In regret theory, one will generally expect r(x, y) > r(x, z) + r(z, y)

with z ∈]y, x[ (rejoice increases more than proportionally in the size

of the difference between bad and good outcome). This means that

r(24, 0) > r(24, 10) + r(10, 0) (condition A).

One will also expect that r(x, y) = −r(y, x) (symmetry between

rejoice and regret) and allows one to rewrite the two equations as

r(24, 0) ≥ 2r(10, 0)

I can accept that the first equation is verified (the feeling of rejoic-

ing when getting 24 rather than 10 may be less than from getting

10 rather than nothing). If that first equation is verified, the sec-

ond equation is more stringent than condition A. However, while

more stringent than usual assumptions of RT, the second equation

does not contradict it in any way. Therefore, such preferences are

consistent with regret theory.

3. We will use the same logic as the one used in 2.b), by making appropriate

substitutions and using dominance arguments.

(First Order Stochastic Dominance)

b) Since E is preferred to C which is preferred to D, then E is preferred

to D (Transitivity)

26 CHOICE UNDER UNCERTAINTY 104

d) C is preferred to D so A is preferred to B (EUT, as in 2.b).

4.

a) The utility of rejecting the lottery is ln(w) (the individual is left with

the same wealth as before), while if the individual plays the lottery,

then she either loses £10, in which case wealth is reduced to w − 10,

or she gains £10, in which case her wealth is w + 11. The expected

utility of playing the lottery is then 21 ln(w − 10) + 12 ln(w + 11), and

rejecting the lottery translates by EUT in 12 ln(w − 10) + 12 ln(w +

11) ≤ ln(w). A straightforward transformation obtains the result.

b) Transforming the inequality in a),

2

(w − 10)(w + 11) ≤ w (94)

w ≤ 110 (95)

2

109

Y ≥ − 109 (97)

9

Y ≥ 1211.111... (98)

2

200

Z≥ − 200 (100)

100

Z ≥ 200 (101)

wealth as much as on the shape of his utility function. Losses that

are big with respect to one’s wealth require high compensation. In

c), the individual requests a very high Y to compensate for the risk

to lose 100. The expected value of the lottery is then required to be

at least £555. It would seem irrational to reject it, but then, one

must put oneself in the situation where one risks to lose 92% of one

wealth, vs. the chance to multiply it by 11. Would you necessarily

accept it?

5.

26 CHOICE UNDER UNCERTAINTY 105

a) Suppose the agent owns the lottery and prices the ticket for the

lottery at pS . The price pS at which the agent will sell must be such

that qu(w + G) + (1 − q)u(w + L) ≤ u(w + pS )

b) Suppose the agent does not own the lottery. The maximum price pB

that the agent would be ready to pay for a ticket to play the lottery

must be such that qu(w − pB + G) + (1 − q)u(w − pB + L) ≥ u(w).

c) I will have pS such that

p

pS = (10 + 10)(10 + 2) − 10

pS = 5.4919

(10 − pB + 10)(10 − pB + 2) = 102

p2B − 32pB + 140 = 0

√ √

which has got two solutions, p1 = 32+2 464 = 26.7 and p2 = 32− 464

2 =

5.2297. Obviously, p1 is not reasonable, so pB = 5.2297.

from EUT for the apparent paradox of the WTA/WTP disparity.

However, in reality, the disparity one finds between WTA and WTP

is much higher than what is obtained here, or that obtained under

any reasonable functional form for u(.), so EUT is not sufficient to

explain the observed disparities.

6.

S1 S2 S3

A a d d

B b b e

C c c c

(p1 , 0, 0, p2 + p3 , 0). B is (0, p1 + p2 , 0, 0, p3 ) and C is (0, 0, 1, 0, 0).

The expected utility of A is u(A) = p1 u(a) + (1 − p1 )u(d). Similarly,

u(B) = (1 − p3 )u(b) + p3 u(e) and u(C) = u(c).

b) Such a pattern of preference is not possible under EUT since it is

intransitive.

26 CHOICE UNDER UNCERTAINTY 106

c) r(x, y) increasing in x means that the higher the good outcome, the

higher the rejoicing.

r(x, y) = −r(y, x) means that regret/rejoice is symmetric. Getting

the good outcome x rather than the bad outcome y produces the

same amount of rejoicing than the amount of regret induced by

getting the symmetric outcome.

r(x, y) > r(x, z) + r(z, y) (condition A) means that the rejoicing

increases more than proportionately with the difference in outcome.

I rejoice more if I gain £100 rather than £0 than if I gain £50 rather

than £0 and then £100 rather than £50, even though the result is

the same.

d) The agent chooses B over A s.t. p1 r(b, a) + p2 r(b, d) + p3 r(e, d) ≥ 0.

The agent chooses C over B s.t. p1 r(c, b) + p2 r(c, b) + p3 r(c, e) ≥ 0

The agent chooses A over C s.t. p1 r(a, c) + p2 r(d, c) + p3 r(d, c) ≥ 0

Summing up the three equations, one obtains:

p1 (r(b, a)+r(c, b)+r(a, c))+p2 (r(b, d)+r(c, b)+r(d, c))+p3 (r(e, d)+r(c, e)+r(d, c)) ≥ 0

are consistent with regret theory, meaning that nothing in regret

theory precludes that this pattern of intransitive choice could happen

(which does not mean it must happen under regret theory!).

e) In the same way as in (d), write the inequalities and sum them

up, and you will see that the sum cannot be more than 0 under

regret theory. This means that while regret theory would allow a

cycle whereby B is preferred to A which is preferred to C which

is preferred to B, it does not allow for the opposite cycle whereby

A is preferred to B, B is preferred to C which is preferred to A.

This type of prediction can then be tested experimentally; verifying

such a pattern would give credence to regret theory if it is shown

that other alternatives to EUT that allow for intransitivity do not

however mandate that cyclical preferences have a specific direction.

27 GAME THEORY 1 107

27 Game theory 1

1.

There are no pure strategy equilibria of this game. (Note: Any provision of

mixed strategy equilibria would be a waste of time and would potentially evi-

dence lack of knowledge of the concept of pure strategy Nash equilibrium).

2.

Z is strictly dominated by X.

b) C is weakly dominated by B. Y is weakly dominated by X. In the

game that remains, A is weakly dominated by B. Knowing player

1 plays B, player 2 plays X. The equilibrium obtained by iterated

elimination of weakly dominated strategies is thus (B, X).

weakly dominated ones, so the game to consider involves only A, B

and C and W, X and Y .

There are two PSNE of this game, (A, W ) and (B, X).

Denoting a, b and c the probabilities for player 1 to play A, B and

C respectively, and w, x and y the probabilities for player 2 to play

W, X and Y respectively, MSNEs must be such that player 1 must

be indifferent between playing A, B or C. I must thus have

w+x+y =1

That equilibrium can hold only if player 2 plays W with probability

1 so one must have:

u2 (W ) = 6a + 2b + 2c ≥ u2 (X) = 4a + 6b + 8c

u2 (W ) = 6a + 2b + 2c ≥ u2 (Y ) = 4a + 4b + 9c

a+b+c=1

a ≥ 0, b ≥ 0, c ≥ 0

(6 − 2b)/8 ≤ a

(7 − 5b)/9 ≤ a

a+b+c=1

a ≥ 0, b ≥ 0, c ≥ 0

27 GAME THEORY 1 108

0, c ≥ 0.

Now, (6 − 2b)/8 ≥ (7 − 5b)/9 when b ≥ 1/11, so we can summarize

by saying that if b ≥ 1/11 then I must have a ≥ (6 − 2b)/8 and

c = 1 − a − b and a ≥ 0, b ≥ 0, c ≥ 0 while if b ≤ 1/11 then I must

have a ≥ (7 − 5b)/9 and c = 1 − a − b and a ≥ 0, b ≥ 0, c ≥ 0.

This is represented in the graph below with b on the horizontal axis

and a on the vertical axis. For b ≤ 1/11, I must have a higher than

(7 − 5b)/9 while for b ≥ 1/11 I must have a higher than (6 − 2b)/8.

I must also have a ≤ 1 − b to ensure that c ≥ 0.

a 1.0

0.9 Range of possible MSNEs

0.8 a=(7-5b)/9

0.7

0.6 a=(6-2b)/8

0.5

0.4

0.3

a=1-b

0.2

0.1

0.0

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

1/11 b

constraints represented above. Note that the PSNE such that a = 1

and w = 1 fits the constraints above.

3.

Left Right

Up 0, 0 4, 1

Down 2, 2 −1, −1

27 GAME THEORY 1 109

b) There are two PSNE in this game, (U, R) and (D, L), and a MSNE

s.t.

u1 (U ) = 4r = u1 (D) = 2l − r

l+r =1

u2 (L) = 2d = u2 (R) = u − d

u+d=1

with solution u = 34 , d = 14 , l = 57 , r = 27

plays U, then 2 plays R so 1 gets 4 while if 1 plays R, then 2 plays

L and 1 gets only 2. Anticipating this, 1 plays U, and this induces 2

to play R.

4.

l r

L 3, 1 0, 0

M 1, 0 0, 1

R 2, 2 2, 2

b)

Nash equilibria:

Note that M is strictly dominated by R, and can thus be eliminated

from consideration.

There are two PSNE in the remaining game, (L, l) and (R, r).

MSNEs must be such that 2 must be indifferent between l and r so

it must be that

p(L) + p(R) = 1

so

p(L) = 0

p(R) = 1

p(l) + p(r) = 1

27 GAME THEORY 1 110

The solutions are thus such that p(l) ≤ 2/3 so the range of MSNEs

is such that p(L) = 0 and p(R) = 1 while p(l) ≤ 2/3.

There is one SPNE such that 1 plays R.

Consider now SPNE such that 1 plays L or M. Denote the beliefs of

2 as µ(R), µ(L), µ(M ), i.e. 2, when it comes its turn to play, believes

1 played R with probability µ(R). Obviously, when it comes for 2 to

play, µ(R) = 0. The expected payoff to 2 of playing l is µ(L), and

of playing r is µ(M ). Therefore,

• If µ(L) > µ(M ), then 2 plays l and then 1 is better off playing

L. By backward induction, under those beliefs, 1 will indeed

play L since this gets it 3 rather than 2 if it played R, so there

is a sustainable belief such that µ(L) = 1 and µ(l) = 1.

• If µ(L) < µ(M ), then 2 plays r and 1 is indifferent between

playing L or M since both get it payoff 0. But then, by back-

ward induction, 1 is better off playing R so the belief that

µ(R) = 0 is not sustainable. We can thus eliminate the possi-

bility that a belief such that µ(R) = 0 and µ(L) < µ(M ) could

be sustained under a SPNE.

• If µ(L) = µ(M ) = 12 then 2 is indifferent between playing l or

r. 1 must be indifferent between playing L or M so I must have

3µ(l) = µ(l), or µ(l) = 0. But if µ(l) = 0, then by backward

induction 1 is better off playing R so the belief that µ(R) = 0

is not sustainable. We can thus eliminate the possibility that a

belief such that µ(R) = 0 and µ(L) = µ(M ) could be sustained

under a SPNE.

There are thus two pure SPNE, one such that 1 plays R, the other

such that 1 plays L and 2 plays l.

c) In the first SPNE, the payoff to 1 is 2. In the second SPNE, its payoff

is 3. By forward induction therefore, the second SPNE will be the

equilibrium of the game. Indeed, when it comes the turn for 2 to

play, it must assume 1 played L since this is the only way it could

get a better payoff than by playing R. Therefore, the only plausible

SPNE, obtained by forward induction, is such that 1 plays L and 2

plays l.

5.

l m r

L 1, 3 1, 2 4, 0

M 4, 0 0, 2 3, 3

R 2, 4 2, 4 2, 4

27 GAME THEORY 1 111

b)

Nash equilibria:

There are no strictly dominated strategies in this game. There is

one PSNE of this game, (R, m).

MSNEs must be such that player 2 is indifferent between l, m and

r. In shorthand, denote pR as R, pM as M and pL as R

3L + 4R = 2L + 2M + 4R

3L + 4R = 3M + 4R

L+M +R=1

into the first, means L = M = 0 and from the third, that R = 1.

In that equilibrium, player 1 must at least prefer playing R, so I

must have

2l + 2m + 2r ≥ l + m + 4r

2l + 2m + 2r ≥ 4l + 3r

l+m+r =1

l ≥ 0, m ≥ 0, r ≥ 0

3m ≥ 2 − 3l

3m ≥ l + 1

l+m+r =1

l ≥ 0, m ≥ 0, r ≥ 0

0, r ≥ 0.

Now, (2 − 3l)/3 ≥ (l + 1)/3 when l ≤ 1/4 so we can summarize

by saying that if l ≤ 1/4 then I must have m ≥ (2 − 3l)/3 and

r = 1 − l − m and l ≥ 0, m ≥ 0, r ≥ 0 while if l ≥ 1/4 then I must

have m ≥ (l + 1)/3 and r = 1 − l − m and l ≥ 0, m ≥ 0, r ≥ 0.

and m on the vertical axis. Note the constraint that I must also

have m ≤ 1 − l to ensure that r ≥ 0.

27 GAME THEORY 1 112

m

1.0

0.9

0.8 Range of possible MSNEs

0.7 m=(l+1)3

m=(2-3l)/3

0.6

0.5

0.4

0.3

m=1-l

0.2

0.1

0.0

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

l

1 plays R is a SPNE. Consider now if there are other SPNE that

involve 1 playing L or R. Denote µ(L), µ(M ) and µ(R) the beliefs of

player 2 once it comes its turn to play, i.e. its estimated probabilities

for 1 to play L, M or R respectively. Obviously, µ(R) = 0 since 2

knows that since it is its turn to play then it must be that 1 did not

play R.

Consider the best strategy for 2 given its belief that L is played

with probability µ(L). The expected payoff of playing l is 3µ(L),

the expected payoff of playing m is 2, and the expected payoff of

playing r is 3µ(M ).

• If 3µ(L) > max(2, 3µ(M )), then 2 plays l, but in that case,

by backward induction, 1 is better off playing M, therefore the

belief of 2 is not sustainable (it should be µ(M ) = 1).

• If 2 > max(3µ(L), 3µ(M )) then 2 plays m, but in that case,

by backward induction, 1 is better off playing R, therefore the

belief of 2 is not sustainable.

• If 3µ(M ) > max(2, 3µ(L)) then 2 plays r, but in that case,

by backward induction, 1 is better off playing L, therefore the

belief of 2 is not sustainable.

• Suppose now µ(L) = 32 so 2 is indifferent between playing l or

m, and does not play r. For this belief to be sustainable, 1 must

27 GAME THEORY 1 113

µ(r) > 0. So, denoting µ(l), µ(m) the beliefs of 1, I must have

µ(r) + µ(l) + µ(m) = 1

µ(r) = 0

so I must have

1

µ(l) =

4

3

µ(m) =

4

2

There is thus a SPNE such that 1 plays L with probability 3

and 2 plays l with probability 14 .

• Suppose now µ(M ) = 23 so 2 is indifferent between playing m

or r, and does not play l. For this belief to be sustainable, 1

must be indifferent between playing L or M and cannot believe

µ(l) > 0. So, denoting µ(m) and µ(r) the beliefs of 1, I must

have

µ(r) + µ(l) + µ(m) = 1

µ(l) = 0

2

There are thus two SPNE: R and a SPNE such that p(L) = 3,

p(M ) = 31 , µ(l) = 14 , µ(m) = 43 .

c) The expected payoff to 1 of playing R is 2 while the expected payoff

of playing the other SPNE is 1, so that by forward induction player

1 should play R. The only plausible SPNE is thus for 1 to play R.

6.

it would be very hard for them to reach an agreement on which one

to play, so they would probably also agree on using a randomizing

device to choose which equilibrium to choose. This randomizing

device should guarantee each player the same payoff in expectation,

so one would need for this randomizing device to lead with equal

probability to one or the other equilibrium being chosen. A coin

toss would do the trick.

b) It would seem reasonable for 2 to play Middle. Indeed, while it could

make more by playing Stop if 1 plays Up, there is too much of a risk

to lose −10 in case 1 plays Down (which may indeed happen if 1 bet

on 2 playing Right).

27 GAME THEORY 1 114

c) There are two PSNE, (Up, Stop) and (Down, Right). The ’reason-

able’ play by 2 evoked above is thus not a PSNE.

d) Denote p(Up) = q.

i) The expected payoff for 2 if playing Left is u2 (L) = q − 5(1 − q),

while the payoff of playing Stop is u2 (S) = 2q − 10(1 − q), the payoff

of playing Right is u2 (R) = −10q +2(1−q) and the payoff of playing

Middle is u2 (M ) = 0.

Left is thus played subject to 6q − 5 ≥ max(12q − 10, 2 − 12q, 0),

that is for q ≥ max((12q − 5)/6, (7 − 12q)/6, 5/6), while I must also

have q = q. There is thus ONE point where L might be played, and

that is when q = 5/6. It is therefore false to say that L will never

be played for any q.

Note: Arguing that Left was never strictly preferred to either of

the other actions got the student full marks, with a bonus for those

exploring the case where q = 65 .

ii) The payoff for 2 of playing Stop is u2 (S) = 2q − 10(1 − q) while

the payoff of playing Right is u2 (R) = −10q + 2(1 − q). Stop is thus

preferred to Right when 2q −10(1−q) > −10q +2(1−q) that is when

q > 21 , and conversely. Now, if q = 12 , 2 ’s payoff of playing R and S

are negative, so player 2 prefers to play Middle (note the payoff for

Left is also negative ). In summary, this means that whenever player

2 might be indifferent between S and R, it actually prefers playing

something else than either S or R, so that in any situation where

either S or R might be played, it must be that the other action is

not preferred.

iii) 2 will play Middle subject to 0 ≥ max(q − 5(1 − q), 2q − 10(1 −

q), −10q + 2(1 − q)) or

5 1

≥q≥

6 6

e) The range where Middle is optimal is large, so that if one assigns

equal probability to q being any number in the interval ( 16 , 56 ) (uni-

form distribution), then the probability to be right in playing Middle

is 23 , which is more than the probability Stop is optimal ( 16 ) or the

probability Right is optimal ( 16 ). Therefore, in the absence of any

reasonable belief over the possible actions of 1, it is best to play M

as a way to reduce the risk of making a mistake.

28 GAME THEORY 2 115

28 Game theory 2

1.

2

L R

a) The game in normal form is as follows:

1 U 1, 3 1, 3

D 0, 0 3, 1

This game has two PSNE, (U, L) and (D, R). Its MSNEs much be

such that 2 is indifferent between L and R, so

3u = 3u + d

u+d=1

U to D, so I need

1 ≥ 3r

l+r =1

1

so there is a range of MSNE such that r ≤ 3 and u = 1.

b) Suppose 1 plays U, then its payoff is 1. Suppose now 1 plays D, then

2 will play R and the payoff to 1 will be 3. Therefore, by backward

induction, the unique SPNE of the game is (D, R).

c) Player 2 can precommit (see end of “Game Theory 2” for the case

where player 2 cannot pre-commit).

i)

28 GAME THEORY 2 116

1,3

U

1

L 0,0

Not pre-commit 2

D

2 R 3,1

1,3

U

Pre-commit 0,0

L

1 2

D

R 3,1-c

ii) Put the game in normal form (not necessary for the resolution, but

helpful). Denote C the decision to pre-commit and N C the decision

not to pre-commit, x the probability with which 2 commits, l the

probability with which she plays l and u the probability with which

1 plays U.

N C, L N C, R C, L C, R

U 1, 3 1, 3 1, 3 1, 3

D 0, 0 3, 1 0, 0 3, 1 − c

There are three PSNE, {U, (N C, L)} and {U, (C, L)} , {D, (N C, R)} .

MSNEs must be such that 2 is indifferent between (N C, L), (N C, R),

(C, L) and (C, R) so I must have

3u = 3u + (1 − u) = 3u = 3u + (1 − c)(1 − u)

For u to indeed be equal to 1, then 1 must at least weakly prefer U

to D so I must have

28 GAME THEORY 2 117

R whenever it comes its turn to play. By backward induction,

this means player 1 will play D whether 2 pre-commits or not.

Knowing this, player 2 does not pre-commit. The SPNE is thus

{D, (N C, R)} .

2. Suppose now 1 − c < 0 and 2 pre-commits. Then the player 2

strictly prefers playing L if player 1 plays D, in which case 1 gets

0. 1 thus prefers playing U, and 2 thus gets 3. Suppose 2 does

not pre-commit. Then, as before, (D, R) is played, and 2 gets

only 1. Therefore, 2 prefers to pre-commit and 1 plays U. The

SPNE is thus {U, (C, L if pre-commit, R if no pre-commit)}, equiv-

alent to {U, (C, L)} .

is the one that is expensive enough that it changes the expected

action of 2 in a credible manner, so it is indeed best for 2 to play

L when it comes it turn to choose between L and R and it has

pre-commited. Note also that the pre-commitment was all the more

effective as it was public (known by all). In the variant in appendix

A, one will examine how important it is for the pre-commitment to

be observable.

2.

between accepting or rejecting the price offer, then it accepts.

1. If p > 60 then the seller does not sell. Profit for the seller is 0.

2. If 60 ≥ p > 55, then the seller sells with probability h and its

expected profit is h(p − 50).

3. If 55 ≥ p > 50 then the seller always sells and its expected

profit is (p − 50).

4. If p ≤ 50 then the seller always sells and its expected profit is

(p − 50)

The seller maximizes its payoff. It will set either p = 60, in which

case profit is 10h, or p = 55, in which case profit is 5. It will choose

one over the later s.t. 10h ≥ 5, or h ≥ 21 .

b)

i)

28 GAME THEORY 2 118

60-p1,p1-50

Buy

Buyer

t(60-p2),p2-50

Buy

High type

Buyer

Don’t buy

Seller sets p1 and p2 N

55-p1,p1-50

Buy

Low type

t(55-p2),p2-50

Buy

Buyer

Buyer

Don’t buy

ii)

1. If the high type accepts in the first period, then its payoff is

60 − p1 > 0.

2. If the high type accepts in the second period, then its payoff is

t(60 − p2 ) > 0

3. The low type does not accept in the first period, as its payoff

would be 55 − p1 < 0.

4. If the low type accepts in the second period, then its payoff is

t(55 − p2 ) > 0.

iii) The conditions for such a pattern of acceptance are such that

60 − p1 ≥ t(60 − p2 )

60 − p1 ≥ 0

55 − p1 < 0

55 − p2 ≥ 0

The expected payoff for the seller is then hp1 + (1 − h)p2 − 50, which

it maximizes in p1 and p2 under the conditions above. This leads

to setting p2 = 55 and p1 = 60 − 5t, so her expected payoff is

h(60 − 5t) + (1 − h)55 − 50 = 5(1 + h − ht).

28 GAME THEORY 2 119

off buyig first, while the price p2 cannot be raised to more than the

buyer of a low type’s willingness to pay, 55, so there is no possibility

of exploiting the knowledge that the buyer who does not buy in the

first period is of a low type.

3.

produce the good than the good is valuable. Similarly, the player of

type c is called efficient because it costs it less to produce the public

good than the public good is valuable.

Contributions must be made at the same time else there will be a

waiting game whereby the players wait for the other to contribute.

Similarly, it is important they cannot observe what the other player

contributes else they could devise strategies such that their contri-

bution is dependent on what the other player contributes. Note this

does not require they contribute in succession: each can simply an-

nounce their menu of contribution as a function of the contribution

of the other and then, having observed each other’s menu, choose

what to contribute.

Those two assumptions therefore considerably simplify the game to

be analyzed.

b) If both players are of a low type, the normal form of the game is as

follows:

C NC

C 1 − c, 1 − c 1 − c, 1

NC 1, 1 − c 0, 0

unique PSNE of the game is (NC,NC).

If both players are of a high type, the normal form of the game is as

follows:

C NC

C 1 − c, 1 − c 1 − c, 1

NC 1, 1 − c 0, 0

There are two PSNE of the game, (N C, C) and (C, N C) and a

MSNE such that, denoting ca the probability a contributes and cb

the probability that b contributes then

1 − c = ca

1 − c = cb

c)

28 GAME THEORY 2 120

(1-c,1-c)

bC

C (1-c,1)

a

NC

C

NC (1,1-c) (1-c,1-C)

b bC

(H,H) C

NC (0,0) (1-c,1)

a NC

(H,L) C

NC (1,1-C)

b

NC (0,0)

N

(1-C,1-c)

bC

C (1-C,1)

(L,H)

NC

a C

(L,L) C (1-C,1-C) NC (1,1-c)

C b b

(1-C,1) NC (0,0)

NC

a C

NC (1,1-C)

b

NC (0,0)

one is of a low type and the other of a high type, then the high

type should contribute. If both are of a high type, only one should

contribute.

e) Suppose sa (ca ) takes the form of a mixed strategy, indicating with

what probability player a contributes depending on its type ca . Sim-

ilarly for sb (cb ). Strategies must be at least weakly optimal for each

players of each different type. Under those conditions, the expected

payoff to player a of high type contributing is 1 − c while the ex-

pected payoff of not contributing is λsb (c) + (1 − λ)sb (c), so for the

player a of a high type to contribute, I must have

s.t.

1 − c ≥ λsb (c) + (1 − λ)sb (c)

However, the second equality cannot be verified since 1 − c < 0, so

I must have sa (c) = 0. By symmetry, sb (c) = 0. This means that

sa (c) = sb (c) = min( 1−c

λ , 1). Note how this generalizes findings in

b).

f) A player of a low type again never contributes as contributing is a

dominated strategy for that type of player. Player b will interpret a

28 GAME THEORY 2 121

type, since only high types contribute. Therefore, and because a will

know b knows it is of a high type, b will not contribute in the second

period if a contributes in the first. This means that presumably,

in the optimal strategy, a contributes with lower probability in the

first period than in the game with only one period in order to avoid

revealing its own type as often. (Note: It was not necessary and not

required to solve the game).

Variation on 1.c):

Let us consider a variant of exercise 1 section c, when player 2 cannot precom-

mit to spend c > 0 when she has to play R : 1 does not know whether 2

precommits or not. This variant will allow us to examine how important

it is for the pre-commitment to be observable.

i)

1,3

U

1

L 0,0

Not pre-commit 2

D

2 R 3,1

1,3

U

Pre-commit 0,0

L

1 2

D

R 3,1-c

As before, there are three (N C, L)} , {U, (C, L)} , {D, (N C, R)} ,

and a range of MSNEs U, l ≥ 23 .

28 GAME THEORY 2 122

whenever it comes its turn to play. By backward induction, this

means player 1 will play D whether it believes 2 pre-committed

or not. Knowing this, player 2 does not pre-commit. The SPNE

is thus {D, (N C, R)} .

2. Suppose now 1−c < 0, then the player 2 strictly prefers playing

L if player 1 plays D and 2 pre-committed, in which case 2 gets

0, while it strictly prefers playing R if player 1 plays D and 2

did not pre-commit, in which case 2 gets 1.

Denote µ the belief of 1 regarding whether 2 did pre-commit.

The expected payoff to 1 of playing D is then 3(1 − µ) while its

expected payoff of playing U is 1. Therefore, if µ < 32 , then 1

plays D while if µ > 23 then 1 plays U, and if µ = 23 then 1 is in-

different between U and D. Note that if one denotes l the prob-

ability 2 plays l, 1 indifferent between U and D translates in l =

2

≤ 23 , L if pre-commit, R if no pre-commit)

3 . The SPNEs are then D, (µ

and U, (µ ≥ 23 , L if pre-commit, R if no pre-commit) .

iii) If 1−c < 0, 2 will precommit s.t. µ > 23 , so for any µ > 23 then 2 pre-

commits and thus the only reasonable belief by forward induction is

µ = 1.

Otherwise, if µ < 32 , then 2 does not pre-commit and thus the only

reasonable belief by forward induction is µ = 0.

Finally, if µ = 32 , then 2 commits with probability 32 , so µ = 2

3 is a

reasonable belief.

There are thus only three reasonable SPNEs obtained by forward

induction when 1 − c < 0, which are:

1. {D, (µ = 0, L if pre-commit, R if no pre-commit)} , equivalent

to {D, (N C, R)}

2. {U, (µ = 1, L if pre-commit, R if no pre-commit)} , equivalent to

{U, (C, L)}

3. U,(µ = 32 , L

if pre-commit, R if no pre-commit) , equivalent

to U, l = 32

depends on the players’ beliefs, while in the original exercise where

precommitment was observed, 2 was able to force the belief of player

1 according to her own interest.

If 1−c > 0, as said above, the SPNE is {D, (µ = 0, L if pre-commit, R if no pre-commit)} ,

equivalent to {D, (N C, R)} .

The possibility to pre-commit thus adds the possibility of a rea-

sonable SPNE that obtains 2 a higher payoff than without pre-

commitment. However, whether it obtains or not depends on what

1 believes 2 believes 1 believes 2 will do.

Note how the process of forward induction allowed one to restrict

the set of NEs and SPNEs.

29 INCENTIVE THEORY 123

29 Incentive theory40

1.

Of the three situations given in the question, case A is the only one where the

real moral hazard problem arises.

• First consider case B. Here Low effort gives the principal an expected

payoff of 0.1 ∗ 10 + 0.9 ∗ 100 = 91, whereas High effort gives 0.9 ∗ 10 + 0.1 ∗

100 = 19. The principal would prefer the agent to give Low effort. The

agent on his part always prefers to give low effort, because it involves low

disutility. So in this case there is no moral hazard.

• Next, consider case C. Here, High effort gives 100 for sure. So if the

principal ever observes an outcome of 10, he will conclude that the agent

must have given Low effort. This helps the principal write a contract in

which an outcome of 10 is always punishable. The punishment can be

made appropriately large so that the agent will be deterred from giving

low effort. So here though the moral hazard problem potentially exists, it

can be solved easily as effort is in effect observable.

• Now consider case A. Here Low effort gives a lower expected payoff (37),

and High effort gives a higher expected payoff (91). The principal prefers

High effort, but the agent prefers Low effort, and the principal cannot

determine the agent’s effort from the outcome observed. Hence, there is a

moral hazard problem.

2.

b) Moral hazard constraint: pG w > B

Participation constraint: pG w > 0

Limited liability constraint is ensured as w > 0

B

Therefore, I need w > pG = 60

haviour, so w = 60. In that case, the return to the investor is

pG (R − w) − K = 90 − 100 = −10. The investor does not invest due

to moral hazard on the part of the manager. The investor may try

to get a manager with some capital to invest in the project, or try to

limit the private benefits B the manager can get when she behaves

badly.

3.

40 Answer keys for exercises 1, 3 and 4 were originally written by Dr Bibhas Saha.

29 INCENTIVE THEORY 124

max πi = Ri − cqi

Subject to Ui = qi [θi − qi /2] − Ri ≥ 0

is qi = θi −c, for i = 1, 2. The monopolist will sell this much quantity

(θ 2 −c2 )

to each type and will charge Ri = i 2 .

R S2

U2

R2*

S1

U1

R1 *

q1* q2* q

at the tangency point of the indifference curve and iso-profit curve

for each type. Note that on the (q, R) plane, higher indifference

curve means lower utility and higher iso-profit curve means higher

profit. Also the indifference curve of type 2 is steeper than that of

type 1. In the figure we have drawn the indifference curves at the

reservation utility level (which must pass through origin).

b) See the graph below:

29 INCENTIVE THEORY 125

R S2

U2

R2*

S1

U1

R1*

q1* q2* q

Figure 29: Outcome of the first best contracts under asymmetric information

in Exercise 3, Incentive Theory

If the monopolist does not have full information and yet it offers

(q1 , R1 ) and (q2 , R2 ) both, both types of consumers will choose

(q1 , R1 ). As evident from the above graph, type 2 consumers’ utility

increases (thick indifference curve below the dotted one), while the

type 1 consumers cannot choose anything else. Hence, the monopo-

list will sell only (q1 , R1 ) bundle and get N (R1 − cq1 ) as profit.

c) The monopolist’s problem is now modified as;

Subject to

Self-selection constraints:

• U1 = q1 [θ1 − q1 /2] − R1 ≥ q2 [θ1 − q2 /2] − R2

• U2 = q2 [θ2 − q2 /2] − R2 ≥ q1 [θ2 − q1 /2] − R1

and participation constraints

• U1 ≥ 0

• U2 ≥ 0.

Of these the second self-selection constraint will bind and the first

participation constraint will bind. This is easy to check. Just assume

otherwise and see that you will run into contradictions.

29 INCENTIVE THEORY 126

efficiency condition will hold for the top type. No rent at the bottom

means that U1 = 0 the reservation utility, which is confirmed at the

binding of the first participation constraint.

Substitute into the objective function the expressions for R1 and R2

obtained by setting the second self-selection constraint and the first

participation constraint with equality, and then maximise it with

respect to q1 and q2 . You should get the following:

• q2 = θ2 − c

and

• q1 = θ1 − c − {(1 − α)/α}[θ2 − θ1 ].

Associated payments (R1 , R2 ) can be easily calculated.

R

S2 U2

R2SB

S1

U1

First best

R1SB

Second best

q1SB q2* q

now fallen as the monopolist’s choice is now given at the point of

intersection of the two solid indifference curves. This follows from

the fact that the self-selection constraint must bind for the top type.

He should be indifferent between choosing his own bundle and the

one meant for the bottom type. Choice of q2 is unaffected, but R2

falls, implying that concession has been given to the top type.

29 INCENTIVE THEORY 127

4.

a) The second firm is the top type. It is the efficient type as it can

supply a given level of cleanliness at a lower cost.

b) Full information solution. The hospital’s problem is to

• Maximise U = AI − P

• Subject to P − kI 2 ≥ 10.

A A2

It is straightforward to obtain: I = 2k and P = 10 + 4k .

Subject to,

Self-selection constraints:

• P1 − I12 ≥ P2 − I22

• P2 − kI22 ≥ P1 − kI12

Participation constraints:

• P1 − I12 ≥ 10

• P2 − kI22 ≥ 10.

straint will bind. By setting equality to these two constraints obtain

expressions for P1 and P2 , substituting them into the objective func-

tion and maximising it, one obtains the following:

• I2 = A/(2k).

e) See the following graphs. In the top panel we have depicted the

full information solution. The bottom panel shows how the optimal

solution will look like under asymmetric information. P2 rises above

the full information solution, while I1 is below its full information

level. If these two bundles {(I1SB , P1SB ), (I2SB , P2SB )} are offered

together, the top type will select (I2SB , P2SB ) as intended.

29 INCENTIVE THEORY 128

P

First best S1

U1 U2

S2

P2*

P1*

I1* I2* I

P

S1

First best

S2 U2

Second best

P2SB>

P2*

P1SB

I1SB<I1* I2* I

5.

29 INCENTIVE THEORY 129

a)

i) The agent that does not exert effort will have expected payoff:

straint ((p1 −p0 )(u(Q+w)−u(w)) ≥ C) does not come into account,

and one only has to check that EUe=1 > 0.

ii) The employer will maximize its objective function EUe=1 (the utility

of the agent) s.t. p1 w + (1 − p1 )w ≥ 0 (this is the budget constraint

for the employer)

s.t. p1 w + (1 − p1 )w ≥ 0

the government maximizes

−p1 w

maxp1 u(Q + w) + (1 − p1 )u( )−C

t 1 − p1

so that at the optimum

−p1 w

u0 (Q + w) = u0 ( )

1 − p1

so

u (Q + w) = u0 (w)

0

that w < 0). This is explained by the wish of the government to

minimize the variability of the income of the agent as the agent is

risk averse. Note that if effort wasn’t verifiable, then the agent would

not do effort as income is the same either way.

b)

i) If effort is not verifiable, then the agent must be induced to choose

to do effort, so the program of the principal is

s.t. p1 w + (1 − p1 )w ≥ 0 (budget constraint)

(p1 − p0 )(u(Q + w) − u(w)) ≥ C (incentive constraint)

29 INCENTIVE THEORY 130

(1 − p1 )u0 (w) − µ(1 − p1 ) − λ(p1 − p0 )u0 (w) = 0

p0

u0 (w) + λ(1 − )u0 (Q + w) = µ

p1

(since u0 (w) ≥ 0 (increasing utility function)).

p1 − p0 0 π0

(1 − λ )u (w) = u0 (w) + λ(1 − )u0 (Q + w)

1 − p1 π1

p 1 − p 0 0 π0

u0 (w) − u0 (w) = λ[ u (w) + (1 − )u0 (Q + w)]

1 − p1 π1

In that case, both λ and µ will be positive at the optimum so the

two constraints, budget and incentive, will be binding:

π1 w + (1 − π1 )w = 0

C − (p1 − p0 )(u(w) − u(w)) = 0

This means that in order to induce effort, the government will have

to differentiate payment between the high and the low payoff in order

to induce the agent to seek to increase the probability of the high

payoff by exerting effort. This comes at the expense of making the

payoff to the agent more uncertain.

6.

a) The principal is the firm as it draws up the contract, while the agent

is the employee. The contract will be drawn based on two variables,

w the wage and h the number of hours of work, and the wage and

number of hours worked may be made dependent on the result of

the shock (positive or negative).

b)

(wH , hH ) is to be paid and worked in case of a favourable shock

and CL = (wL , hL ) is to be paid and worked in case of an

adverse shock.

• At time t = 2, the agent accepts or reject the contract. If the

contract is rejected, the game ends.

29 INCENTIVE THEORY 131

• At time t = 4, transfers are made according to the contract

and the realization of the outcome.

Eπ = q(xH hH − wH ) + (1 − q)(xL hL − wL )

utility EU is positive:

d) The Lagrangian of the optimization problem is

L = q(xH hH −wH )+(1−q)(xL hL −wL )+λ(q(u(wH )−f (hH ))+(1−q)(u(wL )−f (hL )))

Maximization in λ obtains that

Maximization in hH , hL , wH and wL obtains the following first order

conditions:

xH − λf 0 (hH ) = 0

xL − λf 0 (hL ) = 0

−1 + λu0 (wH ) = 0

−1 + λu0 (wL ) = 0

xH

=λ

f 0 (hH )

xL

=λ

f 0 (hL )

λu0 (wH ) = 1

λu0 (wL ) = 1

the same wage irrespective of the shock because the agent is risk

averse. Now, from the two first equalities, I have that f 0x(hHH ) =

xL

f 0 (hL ) so

xL 0

f 0 (hL ) = f (hH )

xH

Since xL < xH this means that

f 0 is increasing as f is convex).

29 INCENTIVE THEORY 132

f) Under the contract above, the principal has to pay the same wage

whether the agent works hH hours or hL hours. This means that in

case of a negative shock, the principal will want to lie and pretend

there was a positive shock so the agent has to work more.

g) One must include incentive constraints for the principal so it does

not lie, so one needs:

xH hH − wH ≥ xH hL − wL

case of a positive shock, and

xL hL − wL ≥ xL hH − wH

case of a negative shock.

Adding up those two conditions, one finds that

xH hH − wH + xL hL − wL ≥ xH hL − wL + xL hH − wH

(xH − xL )hH ≥ (xH − xL )hL

hH ≥ hL

Then, looking at the second incentive constraint:

wH − wL ≥ xL (hH − hL )

case of a negative shock. This means that the first best outcome

cannot be achieved if the principal can lie. The possibility of lying

works to the detriment of the principal.

- StockWatson_3e_EmpiricalExerciseSolutionsЗагружено:markus
- Advanced Microeconomic Theory_Jehle_RenyЗагружено:api-19897963
- Mas-Collel - Microeconomic Theory.pdfЗагружено:Alex Rilie
- Lecture Notes on Microeconomic Theory - Nolan Millers (Havard)Загружено:joydrambles
- Kennedy a Guide to EconometricsЗагружено:Jesús Treviño
- Micro-economic initiatives: handbookЗагружено:International Committee of the Red Cross
- Advanced Microeconomic TheoryЗагружено:api-7805972
- Microeconomic Theory by MasColell & WhinstonЗагружено:morinsola
- Mathematics for EconomistsЗагружено:blakmetal
- Microeconomics of BankingЗагружено:Mike Rodriguez
- Helpman.krugman.1999.Market Structure and Foreign TradeЗагружено:Igor Makarov
- Mas Colell SolutionsЗагружено:Javier Quintana
- Mas Colell Solution ManualЗагружено:usar scribd
- Cyclopedia of EconomicsЗагружено:Sam Vaknin
- Solutions GoldbergerЗагружено:jalexrs
- Microeconomic TheoryЗагружено:sudeepshaw
- Jehle and Reny SolutionsЗагружено:Kevin Andrew
- MicroeconomicsЗагружено:Saule Baurzhan
- Goldberger. a Course in EconometricsЗагружено:Nic Palar
- Mas Colell CopiaЗагружено:mmonturiol
- Advanced Macroeconomics Solutions ManualЗагружено:Jie Yan
- Solution of VarianЗагружено:Md.Shahriar Haque
- microeconomicsЗагружено:Bennie Prince Nesh
- Jehle SolutionsЗагружено:Tamás Szabó
- Stock Watson 3U ExerciseSolutions Chapter9 InstructorsЗагружено:vivo15
- Applied. EconometrcisstatapdfЗагружено:mandcrut
- Solution Exercises 2011 PartIЗагружено:cecch001
- A Guide to Modern Econometrics - Marno VerbeekЗагружено:nickedwinjohnson
- 31077845-Advanced-Microeconomics-Analysis-Worked-Solutions-Jehle-Reny.pdfЗагружено:Benjamin van Arum

- Experimental evidences about the limits of nudges as a tool for public policyЗагружено:Alexia Gaudeul
- Behavioral Game TheoryЗагружено:Alexia Gaudeul
- Consumption, social networking and collaboration on the InternetЗагружено:Alexia Gaudeul
- Fostering Trust in the Advice of ExpertsЗагружено:Alexia Gaudeul
- Why Open Source Software does not SucceedЗагружено:Alexia Gaudeul
- Public and private provision of public and private goodsЗагружено:Alexia Gaudeul
- CFP INFER 2018 GöttingenЗагружено:Alexia Gaudeul
- Poster INFER with wordsЗагружено:Alexia Gaudeul
- Do consumers prefer offers that are easy to compare?Загружено:Alexia Gaudeul
- Do Consumers Prefer Offers that are Easy to Compare?Загружено:Alexia Gaudeul
- Poster INFER with picturesЗагружено:Alexia Gaudeul
- Do Open Source Developers Respond to Competition? The (La)TeX Case StudyЗагружено:Alexia Gaudeul
- Open source software and innovationЗагружено:Alexia Gaudeul
- Partnerships, Imperfect Monitoring and Outside Options: Theory and Experimental EvidenceЗагружено:Alexia Gaudeul
- Privacy Protection and Risk AttitudesЗагружено:Alexia Gaudeul
- BA Seminar in Experimental EconomicsЗагружено:Alexia Gaudeul
- Spurious complexity and common standards in markets for consumer goodsЗагружено:Alexia Gaudeul
- The role of reciprocation in the formation of social networks, with an application to bloggingЗагружено:Alexia Gaudeul
- “Many a slip between the cup and the lip”: The effects of default-based nudges on pro-social behavior and attitudes.Загружено:Alexia Gaudeul
- Some Industrial Economics of Software Marketing and DevelopmentЗагружено:Alexia Gaudeul
- Topics in Economic AnalysisЗагружено:Alexia Gaudeul
- Presentation at the 13th Annual Conference of the APET in Taiwan, 2012Загружено:Alexia Gaudeul
- Robustness of the attraction effect in consumer choice and its consequences for the firmsЗагружено:Alexia Gaudeul
- Of the stability of partnerships when individuals have outside optionsЗагружено:Alexia Gaudeul
- Consumer theory: Preferences, consumption and demand.Загружено:Alexia Gaudeul
- Blogs and the economics of reciprocal (in)attentionЗагружено:Alexia Gaudeul
- Do common standards promote competition? A market experiment.Загружено:Alexia Gaudeul
- Why is shopping such a baffling ordeal, and what can you do about it?Загружено:Alexia Gaudeul
- Call for Papers - Fourth FLOSS WorkshopЗагружено:Alexia Gaudeul

- Panasonic Imaging Corp. vs CIRЗагружено:brendamanganaan
- aylawing resumeЗагружено:api-403029141
- 5636015 Solaris Admin Interview Questions AnswersЗагружено:ajin_ss3576
- scm1.pptЗагружено:Saptaparna Ray
- TFT HistoryЗагружено:Philip Wren
- CCP102Загружено:api-3849444
- 15 Road Inventory Survey Report ExampleЗагружено:AJ Sa
- Standing OrderЗагружено:Anoop Chaudhary
- A Qualitative Examination of Barriers for Urban and Rural Custodial GrandparentsЗагружено:windaRQ96
- A Criminal Conspiracy Exists When Two or More People Agree to Commit Almost Any Unlawful ActЗагружено:jemima
- Tempting Failure Study Boxes (2014)Загружено:info8406
- MKT 382 Strategic Marketing Mackie.docx.docxЗагружено:stratetrade
- Targeting PoorestЗагружено:Dxtr Medina
- Ethics and Psychology: Dilemma 6: Referral and Treatment BoundariesЗагружено:Sarah M.
- 1MRK502033-UEN_-_en_Application_manual__Generator_protection_REG650__IEC.pdfЗагружено:Muhammad Syaifulhaq
- Louvre MuseumЗагружено:Amit Rakshe
- New ResepЗагружено:Agatha Tanti
- m3uЗагружено:Anonymous TiDEl9
- A Glossary of Computer-related words in TamilЗагружено:Shalini Kalyani
- Acid ELECTROforesisЗагружено:Yusuf Indra Sentosa
- AB AM 2000Загружено:Ja Six
- List of Books and JournalsЗагружено:dentistdentistdentis
- Research - Chapter 4Загружено:gerard_xucn
- DescartesЗагружено:Pablo Olivares
- Canon_IR_Advance_8205-8285-8295Загружено:copyhere
- musicmagickangels.pdfЗагружено:alxberry
- 2411114 Bdts Chbk Tat 12471 27-11-2011 Zeenat a Gani (Himachal Pradesh)Загружено:zaheerbc
- american-jurisprudence-on-business-trust.pdfЗагружено:HeyYo
- Wakefield Press - October 2017Загружено:Artdata
- Using Iserver Integration TechЗагружено:ashish09390